While still possible, the proof exploiting a bug is vanishingly small now. Following a few bugs and bad proofs this year, there are a number of different implementations, in different languages, and with different levels of proofs of their consistency. con-leche and con-ron are the heavy hitters, with consistency proofs of themselves. So for RH to be false, there'd have to be a bug in all the kernels that affects both the consistency proofs and the RH proof itself.
Which is perhaps not such a far-fetched possibility?
While I have not studied this topic enough to really understand _how_, but I've been told that type systems laid out in frequently cited papers by the likes of Girard and Martin-Löf were found to be unsound more than a decade after publication. These issues were then "patched", so to speak.
Can anyone better versed in type theory explain how likely that is for the type system of Lean?
Con leche has been proved to be consistent if zfc is, so if it turns out to be inconsistent the entire foundation of modern day math turns out to be broken
how unreasonable is it to think that during the ai assited claude bruteforce process, it bruteforced also the kernel and built on a randomly found logical inconsistency
Which isn't actually unreasonable. The chance of it happening is vanishing small, but collective errors have existed before in software and humans (usually stemming from some shared misunderstanding)
FTR I am not saying this is an example, just that it's always possible.
I keep seeing people referring to bruteforcing, and I don't see how the concept makes any sense at all in this context. If one was to "bruteforce" a lean proof, why would that require AI at all?
In empirical research you would call that p-value hacking. The improbable becomes more probable. However, IMHO, randomly hitting a bug seems still quite improbable but certainly more probable as if this was a single human. LLMs rather being optimistic than doubtful certainly does not help in this case. But I guess like in other software development running other agents to find those bits in the proof seems equally feasible.
That is not what bruteforcing means. Parallel work is basically how modern science operates - there is approximately 10 million active human scientists across science, which means thousands studying the same thing globally.
For an analogy ... imagine 10000 lockpickers studying the same type of lock and trying to unlock it. That is what it is.
Bruteforcing would mean smashing the lock with a sledgehammer. You don't even need 10000 sledgehammers for that. The difference is that you don't try to understand the lock, you just overwhelm its construction by sheer force.
That is basically what brute forcing is in computer science.
Imagine I'm try to guess a key design that works and spin up an insane amount of compute to keep trying different designs and eventually find the needle in the haystack. That's basically bruteforcing it because you find it due to compute power instead of any particularly elegant solutions.
Because LLMs are trained on a massive corpus, require hundreds of billions of dollars of resources, have billions of parameters and run on manhattan-sized datacenters.
So yes it's not the traditional definition of bruteforce, but yeah. LLMs are brute forcing stuff
LLMs indeed wield huge cognitive force, bigger than our brains can, but being really smart is not the same as bruteforcing stuff. If it was, we would call Einstein or von Neumann bruteforce scientists. They weren't, they were just very, very smart.
It's a vague bit of terminology. Perhaps "guided fuzzing" might be a better term (such fuzzing still often being considered a kind of brute forcing). You're still throwing a ton of stuff at the wall and seeing what sticks but you're been much smarter about it than just generating completely random strings and checking each of them. Generally it's intractable to generate a Lean proof by 'pure' brute force: the space of possible proofs quickly dwarfs the universe, so you are really depending on the intelligence to do most of the work: the raw number of attempts is what pushes something that's currently barely possible over the finish line.
(Also, it's not obvious how much raw compute was spent on this: the claim of the proof mentions fairly close engagement as opposed to OpenAI's 'massive unattended swarm' approach to the Navier-Stoke's problem. Obviously quite a lot would have been available but I don't think there's public information on how this was actually done)
You just described all of research which uses the scientific method which when it comes down to it is making educated guesses (fine, hypotheses) and then validating them in some logical or empirical way.
Well yeah, if we can find the bug(s). The big worry is that we might overlook the bug and go on our merry way assuming the claimed results are true when they're not. (I don't see that happening in the case of the Riemann hypothesis, but the same worry exists for all proofs that exist only in the form of Lean code, such as the massive dump from OpenAI.)
I wouldn’t be so sure - people have assumed both the Riemann hypothesis and its negation for 150 years and haven’t been able to derive a contraction in either case.
What would be interesting is if we challenged an LLM to prove a known false statement close to the Riemann hypothesis. If it manages to do that, we know it’s exploiting a bug.
That would be interesting indeed, but it would be difficult to control for the circumstance that the LLM likely knows that the statement in question is known to be false (while it certainly knew Riemann to be an unsolved problem).
While it's possible there is a bug, the models are better at finding bugs than they are at proving theorems. I find it hard to believe a model can deliberately exploit a bug to obtain a proof, but the same model cannot find and demonstrate the bug if tasked to do so. That leaves the option that there is a bug that accidentally let us incorrectly prove GRH, without having any know impact on any other proofs, and having exactly the same effect in all lean implementations. It seems too unlikely to be a reasonable concern at this point, except perhaps for the people responsible for administering the price money.
I agree, but I expect the models to reuse the bugs to prove any crap. Like:
fake quote> In section 1 we prove that the RH, in section 2 we prove ¬RH, and in section 3 we use the previous results to prove that 2+2=5.
It is possible that the bugs are so subtle that are only useful to make a fake proof RH but not other false stuff.
But from my experience adding bugs to Racket (and cleaning them mostly before sending the PR or before the release), once you find a bug, it's posible to minimize the code, isolate the bug and create a regresion test that is almost comical because it's obviously wrong.
Since I'm not terribly familiar with Lean, math folks are generally satisfied the provided proof encodes the RH correctly I presume?
edit: I need to learn to read, that's the first part of their statement.
The statement of the Riemann Hypothesis is correct. This statement is part of Mathlib: RiemannHypothesis and has been carefully vetted by experts in Lean and number theory.
Note that this definition is years old. It's very well vetted. It's not a case of "the proof works, it just doesn't actually prove what you think it proves".
If you read the blog post, it says they added a "Challenge.lean" file. What that means is that instead of taking for granted the proof is trying to prove Riemann, they added a small file which reads RiemannHypothesis from the Mathlib library mentioned above, and forces the proof to match that.
They also mention running comparator, which is important because it checks for all the ways a proof might mathematically "cheat". For example, if you are trying to prove 1+2=5, and you redefine 2 as 4, then your bad proof would actually compile in Lean at first, but then comparator checks would discover all the tricks like that and declare the proof invalid.
Additionally, they mention compiling the proof with different Lean kernels, not just the mainline one. This rules out a way a proof This means the proof cannot just take advantage of a bug in the mainline Lean kernel. Importantly, they mention compiling the proof with con-leche, which uses a kernel that has been mathematically verified. Using con-leche is a lot slower, but the fact that the proof works on multiple kernels including con-leche means that it's not taking advantage of some bug in Lean.
At this point, I would say the proof is almost certainly true. The only obvious way it could be false is if ZFC itself is inconsistent or something wild like that, and the proof took advantage of it. But if ZFC is inconsistent, all of mathematics is in trouble.
Given such a massive proof, is there any possibility of accidental tautology inadvertently hiding inside the mess to the effect of "Assume A is true. A million lines later: therefore A is true"?
Lean keeps track of any such assumptions when constructing the proof term, and would not allow them to line up with an unconditional statement like RH.
Of course it's possible that the proof is flawed, in that or numerous other ways. That's why it's being formally verified in multiple independent ways.
That there are tautology's in the billions of line? Yes, absolutely. There's every chance that the proof isn't efficient and includes lines that can be removed.
That the end result relies on those tautologies? No. All lean would derive from doing that is "a implies a" (which it would let you introduce directly) and not the statement "a". This is what computers are perfect at, keeping millions of different records straight and not accidentally forgetting the first half of the implication.
I think they intended to ask about whether the proof could effectively introduce unproven statements equivalent to RH as axioms (i.e. unintentionally proving "RH implies RH" rather than just "RH" by assuming RH at some point within the body of the proof; or the same, but instead of assuming RH, assuming some other statement necessary to the proof).
I believe the blog post does answer that question, though. They say in the article above that "The four criteria in https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i... are all satisfied", and Question 3 in that linked post covers such cases. ("How do we know for sure that there is no trick hidden in the millions of lines of proof files which [some words omitted]... adds a new axiom to mathematics saying that the theorem is true?")
True but irrelevant ... in fact, that's a fallacy of affirmation of the consequent. That every statement can be proven from logical explosion (if ZFC is not consistent) does not mean that every proof is a result of logical explosion.
Consider Euclid's proof that there is no greatest prime. It is valid regardless of whether ZFC is consistent.
Similarly, the Lean proof of the Riemann Hypothesis depends on Lean's type theory, not ZFC's consistency. It's an explicit machine-checked proof term, not a non-constructive consequence of logical explosion.
> the Lean proof of the Riemann Hypothesis depends on Lean's type theory, not ZFC's consistency.
I believe it depends on Lean's type theory which itself depends on ZFC and that if ZFC were inconsistent then Lean's type theory would also be inconsistent.
> It's an explicit machine-checked proof term, not a non-constructive consequence of logical explosion.
It is possible to make "constructive proofs" that make sure never to use an argument saying "not-foo is false, therefore foo must be true". Such a proof really would survive even if ZFC were inconsistent (at least inconsistent in certain ways). However, I don't believe this proof of the Reimann Hypothesis is constructive. (I haven't read the proof, but if it were, that would be remarkable and people would be discussing it: most proofs are not constructive.)
People seem to say that Lean's type theory is strictly stronger (in consistency strength) than ZFC and assuming the existence of any finite number of inaccessible cardinals and strictly weaker than ZFC and assuming the existence of a countably infinite number of inaccessible cardinals.
Isn’t Lean stronger than ZFC? I thought you could prove Con(ZFC) within Lean. So it’s possible that the formal system encoded by Lean is inconsistent but ZFC is not. (I don’t consider this that likely, though.)
isn't ZFC also consistent with ZFC + infinitely many inaccessible cardinals (I assume so or you wouldn't have even said it) which means by the same reasoning ZFC is stronger than ZFC?
Yes, Lean is stronger than ZFC. But two of the checkers, con=leche and con-ron, don't just verify based on Lean. They come with a formal proof from set theory that they cannot wind up accepting an incorrect proof.
In other words you can ignore Lean. We have a proof from ZFC, of the Riemann Hypothesis. We have a proof, acceptable in ZFC, that this proof is correct.
In other words Lean's dependent type theory is used to arrange its understanding of the proof. But the actual reasoning that you wind up with for this proof does not need anything that doesn't follow from ZFC.
why is the verification is so fast this time while last time (a few weeks ago? and I can't remember name of that theorem) it took a few weeks and they finally found an error in Claude's proof?
I am not certain which "last time" you are thinking of, but I suspect that the answer IS the Lean verification.
If an AI writes a paper with an argument that X is true, it might be that X is true, or it might be that the argument in the paper is subtly flawed and X isn't true
On the other hand, if an AI generates a step-by-step proof in which each step can be mechanically verified, then people might have problems understanding the underlying ideas, but once it has been verified we know it won't have subtle errors. And that is basically what Lean is.
In this case, they not only have a Lean proof, they've verified it with more than one verification tool (so it's unlikely to be a bug in the verification tool) and double-checked whether the thing being proved really is what we intended to prove. That's rock solid.
Getting an UNDERSTANDING of the proof is another challenge, but at this point we can be quite certain it doesn't have errors.
Navier-Stokes that is. Claimed proof in September, found an error in October, mainly that the formal language proof does not correspond to the natural language description. But this time it took one day. that's 30x speedup not explained.
The guy waited until he has the verification before anouncing it. I'll quote myself from the pararell universe were he announced it with a English only proof:
This time the WHOLE proof formalized in lean, but iirc some recent AI generated proofs only had supplemental pieces lean that didn't cover the whole argument
I wonder if in N years, when we (might) have crazy ASIs that make Fable 5.2 look like a nematode, maybe we'll see things like this as completely trivial and banal.
A machine intelligence might devise something as (seemingly) difficult as the Riemann hypothesis, but checking if what it posits is correct might be like a fast test command in a git repo, and so it can keep iterating on many ideas without being bottlenecked by waiting centuries or, perhaps even hours, to know if the ideas are true.
But then one could even imagine they'll one day be so powerful they could do this and more without ever using a formal system. They will just naturally be that comprehensive and rigorous internally, with something like Lean making them 100% instead of 99.99..% confident.
And eventually this would extend to almost everything, besides maybe biology where it may only be able to compress a human century into a day rather than into a second.
I think caring about things like the riemann hypothesis is something humans from the before times did. Once we have ASI, there is no incentive to push the bounds of knowledge, or even to try.
I think we're looking at the long dark teatime of the soul of biological human science, mathematics, and engineering.
I guess we will maybe be happier making art and being pets, best case scenario. Or maybe those of us with the most resources will be able to merge into a machine substrate and continue to contribute in some way.
I think it's because you no longer need to release knowledge if everyone can re-derive it independently. Work and learning lose its value when anyone can reproduce it without you, and when they can do it more trivially even than searching for your version. We're starting to reach the point where certain types of projects are just pointless now, because anyone who wants it can make their own version even better tailored to themselves. This isn't absolutely true of course, but the general discourse is heading in this direction, and this is one central focus of many modern imposter syndromes.
So everyone left without these tools to themselves is left in the dust. There are no more publishings of the most advanced knowledge and learnings, because everyone is expected to re-derive it from the prerequisites. But there's so much to derive that you just can't do it all on your own. It's not as natural and effortless as a large language model working through the course of whatever may require the knowledge and deriving it all in a depth first manner.
Mathematics could stall in some ways if the AIs continue to use brute force to prove these. Normally a mathematician would invent powerful new frameworks to solve a difficult problem (like Newton and calculus - there are many recent examples, as well).
As an outsider to the field, the strongest motivation for top mathematicians seems to be pride. I think the educational pipeline for top-level math will weaken if AIs are getting the credit for mathematical work.
This is just a performance issue. In theory, the Lean kernel could be written in Lean.
These independent kernels do not prove their own consistency -- rather, their function is to act as slower but more idiomatic/simpler implementations of Lean's meta-theory. It doesn't matter that they don't correspond 1:1 with Lean's official kernel at either a theoretical or an implementation level, because the independent kernels are not for proving Lean correct -- they're for auditing complex proofs, in a simpler system.
What is particularly interesting is that then these independent kernels are "proven correct"! This notion of correctness is specifically that con-leche (et al) will only accept theorems that are also correct with regards to a particular model of set theory, with membership / extensionality / other standard ZF axioms, and an infinite hierarchy of universes. This is (vaguely speaking) a set-theoretic model of Lean. It is here that Gödel's incompleteness theorems come into play. Lean indeed cannot verify that this set-theoretic model is "correct": what it can verify, however, is that the type-theoretic machinery is correct with respect to the set-theoretic machinery. This removes the possibility of an invalid proof slipping by by way Lean's comparator (which can run alternative kernels) by way of an implementation bug.
It leaves the possibility that Lean's metatheory is inconsistent. This is quite unlikely. The foundations of mathematics have been well-studied, and Lean builds directly on the huge amount of work that has been done, particularly by the Coq community, in figuring out what is and isn't consistent. The Coq theorem prover was really developed in conjunction with modern type theory... Lean came in after a lot of meta-theoretic issues had been ironed out, thanks to them.
This is incorrect. There are multiple kernels that are written in Lean: con-leche (written in Lean) and lean4lean (written in Lean 4). They both agreed that the proof is sound. See https://math.mit.edu/~drew/grh/ for details.
They can't — any kind of formal verification is always with respect to a TCB — but you can get pretty high assurance through making the TCBs smaller and smaller, and making multiple independent TCBs. The Lean community has done both. I think at this point the chance of there being an error is smaller than the chance that aliens show up in the next year or similar.
Imagine 300 millions ago, something living on a not far planet, thinking, is there anybody out there? Doesn't know that space and time are entwined meaning it's not a question of where, but a question of when
When the same bugs show up in multiple "independent" implementations of the "trusted computing base", I start suspecting that they're all ports of each other, making the same unforced choices of algorithm etc. Many folk from the LCF school will point and laugh at the description of Lean's TCB as "smaller". (Does the Lean kernel still contain an entire copy of GMP?) If reimplementing the kernel is much more than a weekend project, I'm certainly not willing to bet against the existence of errors, regardless of whether people claim to have proven their absence.
The "Summer of Soundness Bugs" was the main motivation for the development of con-leche. https://leodemoura.github.io/blog/2026-10-15-lean-now-ships-... con-leche makes this pretty unlikely to occur again, though. It's a formally verified kernel with respect to a set-theoretical model of Lean, that's been implemented in both Lean and Rust, and also has a version that's been compiled using a formally verified compiler with regards to the Kraken x86 semantics. The set-theoretic model theoretically could be inconsistent, but it's quite reasonably designed (ZFC w/o infinity or choice and w/ a universe hierarchy), so that's unlikely.
That line is not correct -- likely AI slop. I have no idea why he does this, but Leo de Moura appears to slop out about half of his blog posts. Not like, half of his overall blog posts, but half of each blog post -- every one is this weird mixture between AI prose and his own prose, from what I can tell from the tone (and the factual issues, like this one). It's really unpleasant. Sadly, they're also canonical reports, as far as the history of Lean goes.
The same bug (in the handling of nested inductive types) was indeed present in two kernels. It was present in Lean's standard kernel, and it was present in the independent lean4lean kernel. The "Collatz proof" using it actually slipped by three kernels in total, because the construction it used just-so-happened to also exploit a different bug in nomeda.
No. Those aspirations are not possible. Mathematics cannot be reduced to pure logic alone - you have to assume something.
However we are now extremely clear on what you need to assume. And why some assumptions are necessary. And we have collections of assumptions that are very reasonable indeed.
The result is that there is a step of faith. But it is a step that people are very comfortable and confident with.
That said, I would be fascinated to know if this proof can be transformed into a proof that does not need all of the power of Lean. Can it be turned into a proof using the Peano Axioms alone? I think that has a good shot of being true. But it would take a lot of work.
We have many proofs given a set of axioms. But those axioms aren’t even taken on faith as much as they are assumed to be true. In a universe where they are true, the conclusions are true. We could assume they are false and prove other things about universes for which they are false.
I’m thinking along the lines of Euclidean vs. non-Euclidean geometry. Euclid made certain assumptions about geometry that were instrumental in proving various theorems. But you can also find geometries where those axioms are not true! And the existence of other geometries doesn’t invalidate Euclid’s conclusions, it just means that they’re only true when its axioms are taken.
Assuming we're not going down the route of paraconsistent logics (which ZFC etc. are not), consistency isn't really something you can do anything interesting with if it's violated. If it turns out you can also prove the Riemann Hypothesis false under ZFC (which would take out every other theorem along with it), that's just bad for math. So we have to take the leap of faith that won't happen.
The core set of things that, if you assume they are correct, the proof holds.
For a non-Lean example, a memory-safe language like Rust has unsafe code that forms its TCB. Assuming that the unsafe code is correct, safe Rust can't cause undefined behavior.
To be clear, formal verification is both with respect to a trusted computing base, and with respect to a theoretical framework. Any such guarantees from formal verification are contingent on the theorem statement being correct, the implementation of whatever framework you're using being trustable, and the meta-theory being sound.
Rust doesn't actually have a theoretical framework, as far as I'm aware. I don't think that existing work (ex. RustBelt) have come close to capturing the full extent of Rust in practice yet. As such, it's totally possible in principle for safe Rust to have theoretically unresolvable memory safety bugs... although it's pretty obvious that it like, doesn't.
I am sure you will get some more detailed answers; I am a bit lazy today; so I will give you the very high level / informal version of the answer (Edit from the future: It grew longer than expected, whelp :D):
Gödel's two incompleteness theorems basically amount to "As soon as your system can do a ridicolously little amount of things, then either of two things is true in this system: 1) You can express facts in your system that you cannot prove within your system or 2) your system claims that some false statement is true."
Now, because 2) is way way worse than 1), you design your systems so that you are firmly in the 1) camp - that way, we gain full confidence in the results our systems proofs, and the price we pay is that our system is less powerfull than it could be.
With Lean, it's similar. The foundation of Lean comes from type theory. There, we do something that is familar to above: We restrict our power (i.e. "which Lean statements do get verified by lean?", and we gain absolute confidence in its proofs (ignoring implementation errors, hardware limitations etc. for now). Intuitively: We have to work harder to satisfy the typechecker/prover, but when it is satisfied, then we have confidence in it.
-------------- Deeper down the rabbit hole
You might have noticed that I seem to muddle the waters a bit with programs, proofs, types and so on... There is a deep reason behind it! In a very strong sense, programs = types = proofs... So, here is how it is all connected (including the RH):
- Lean is a programming language (like Haskell, Rust, ... There is nothing stopping you from writing a server in Lean and deploying that binary somewhere in the cloud. Regarding developer experience, you could think of it like a modern Haskell without lazy evaluation and without its historical warts)
- programs can be tied to types: We first check if the whole program as such has a type ("do all of its parts fit together in a strictly defined way?"). Only if it does, we proceed (and for example, run it). If it does not, then the compiler declares "this program is not good! here is a type error!". Now, here is the price hat Gödel proved we cannot avoid (if we want to trust the typechecker): There are programs that don't type check, but behave perfectly fine at runtime. (That's the "we limit our power" part - those are perfectly good programs where the compiler throws their hand in the air and gives up!)
- Enter the third puzzle piece: proofs and the Riemann Hypothesis: turns out, we can think of types as theorems, and a successful type check as a correct proof! And the RH is a hypothesis we would like to prove...
- So here is how it all ties together: RH is written so that it can be used as a type, in the Lean programming language. The "main program" is basically just another ordinary function with some input types, and returns the type version of the RH being true. Because the function needs an implementation, we are forced to implement that function body - which are many small functions that correspond to mathematical theorems/lemmas if you look at it with the proof glasses on. From the programming pov, it is just functions that call functions that call functions and so on. And from the typechecking point of view; we compose a bunch of functions and check whether input types and output types match up.
Finally, the payoff: *If that program typechecks, then we know RH is true*, which is what just happened!
(PS: The same principle is behind the computational proof of the four-color-theorem. The proof of the four color theorem is "simply" a program that typechecks, where the final type happens to correspond to "can we color any planar graph with at most 4 colors?")
-------------- Further reading
If you want to read up more about this, here would be some sources (note: I am not up to date with how the cool kids these days learn about this topic; so there might now be better resources to learn from!):
- Here is how I learned about it first: https://existentialtype.wordpress.com/2011/03/27/the-holy-trinity/
- wiki: https://en.wikipedia.org/wiki/Curry%E2%80%93Howard_correspondence
- "That's all cute and stuff, but basically fast food for me, give me the good stuff!" okay, okay! how about...
- tying quantum physics, topology, logic and computing together? https://math.ucr.edu/home/baez/rosetta.pdf
- anything on the nLab (a wiki/community(?) that collects and explores those kinds of things... nerds, I tell ya (no, I am not envious at all, naaaah)). For example, this page is about the relation between logic and types: https://ncatlab.org/nlab/show/propositions+as+types
So while I'm not going to question if that's good approach, because if it's so popular, it must be good enough.
But on intuitive level it looks so weird to encode math in types.
When I was taught math, the proofs were more like sequence of statements. So if I would decide to implement proof checker, I'd invent formal language, which looks very close to commonly used math language and which allows to just write proof like it's written in book. Probably a lot more verbose, without skips and implied statements, but in the end, it'd look just like a verbose and formal proof.
Why did math gone this lean type way and not the way I described?
LEAN code looks actually exactly like you propose; it supports most conventional (Unicode-linearizable) math notation, and you just write expressions like in any other functional language. Close to the textbook, but more verbose and stiff as it's a formal language, while math notation isn't.
That lack of relative verbosity and stiffness of mathnotation (what a thing to say haha) to be considered as a proper strength evtly. I'm consider myself as layity in this, but is an interface thinkable?
This is a very good question! There are two answers.
First off, you can absolutely do what you are describing. There are systems like Metamath that assume nothing beyond first-order logic and the axioms of set theory, and apply these verbose rewrites. See https://en.wikipedia.org/wiki/Metamath for more. However, Lean (and many past theorem provers, too) have 1) preferred to represent mathematical objects as types rather than sets, and 2) built a large amount of interactive tooling rather than the blind rewriting of Metamath. The first choice is essentially of elegance -- the Curry-Howard correspondence indeed lets you write propositions as types and proofs as terms inhabiting those types, which is a connection of deep beauty. Ontologically, there is also an argument to be made that representing a mathematical object as a type is more "correct". We don't really think of ex. two as "{{},{{}}}" or "{{{}}}". We think of it as "Two", or maybe "Zero +1 +1". This is a win for types. Mathematical objects are not grocery bags.
Second, the historical approach to theorem proving has been interactive theorem proving. Theorem provers were supposed to be a second brain for students and computer scientists and mathematicians... keeping track of assumptions, suggesting possible tactics to take... aiding the human in the process of proof (and subsequently understanding), rather than replacing them. The path of uncritical adoption of AI that Lean has taken, sadly, much replaces the human. There is no real pragmatic reason for Lean to retain its interactive (human) elements any more (besides inertia) as AI scales up. It might as well be Metamath, indeed.
To add to the other answers, let me address the experience of "a series of steps" versus "a big single expression with a type" - real life is calling, so I have to make this quicker than I would like to; feel free to ask any followup questions / for clarification! I am sure others or later, I when I am back on the screen can answer :)
So the very short version: There is no difference between those two worlds if you squint your eyes hard enough! They are equally powerfull; just different ways of expressing the same truth. This is the whole declarative ("WHAT should the end result be") vs. implerative ("HOW to achieve the end result")stuff, but it's all the same at the end. Some examples:
- SQL is declarative; but under the hood, it will get translated to a sequence of operations on the database by the query engine (or however it's called)
- HTML is declarative; but there is no theoretical reason stopping you from building some sort of "imperative" HTML where you just add elements and stuff in a sequence of steps and at the end; are shown the same webpage (in fact, there are GUI frameworks that build up the GUI in an imperative manner; for the web, transmitting a declaration of what you want rather than a program that constructs it was the more practical choice I assume)
- Haskell is often considered declarative (for good reasons), but if you think about the IO type as ("Hey I am a wrapper type and all I do is to enforce that step A happens before step B"), then you are pretty close to a imperative way of doing things again
- Go/Java/C#/Python with types are almost always considered imperative; but you can also put on the type checking hat and pretend that the program is just one big function that will - if written correctly - have a final type. How does composition happen here? It's basically the line break, line :103 is executed before line :104, so lines :103-:104 can be viewed as: "the function that does whatever :103 does, then whatever :104 does", and it has the type "whatever input type :103 has and whatever input type :104 has", and type checking ensures that all those pieces fit together neatly.
They can even be mixed; in the proof assistant Rocq (formerly known as coq), there where both styles used often. In Lean, as I vaguely remember, you can pick both style too, so to say.
I think the pragmatic answer is that most such languages originated from the "one big function" school, and so it is more natural to program/proove in that way.
---------------------------
*Wait, does this mean that every proof I learned can also be done in this "one big function" way, instead of "a series of steps"?*
Yep! But you will find that it is not really practical to do so. A good exercise for sure though; and something I did (by that I mean: infering types on paper; not really taking proofs I know and doing them be type inference on paper) by hand when diving deeper into this topic in grad school.
The main issue is on how you define a "Type". There are two different logics here;
1) Classical Logic/Set Theory (this is the one you are talking about). Here Type = Set of values defined by membership rules (extensional/intensional).
2) Intuitionistic Constructive Logic/Type Theory (this is what is preferred for proof assistants). Here Type = Set of constructive (often inductive) proofs where the proof itself is the value.
As an example, think of defining the Type "Even_t"; under (1) it can be a predicate "x % 2 == 0", while under (2) it is defined inductively as "0 is Even_t; if construct(n) is Even_t so is construct(n+2)".
Notice that in the latter case a Type maps nicely into the idea of "a computation" and the fact that the very act of generation of one of its inhabitants constitutes "the proof" that the value has that property!
Whereas in Set theory Types are just subsets. But because you can mix anything into a set, an untyped term can be nonsense and will only get caught when you try to use it. Type Theory disallows it by construction.
This idea of "Typing" can further be extended (with additions of Dependent Types etc.) giving a complex matrix of powerful Type Theories represented by what is known as "Lambda Cube" (https://en.wikipedia.org/wiki/Lambda_cube) and their implementations in proof assistants eg. Calculus of Constructions (CoC), Calculus of Inductive Constructions (CIC) etc.
Finally, note that not all proof assistants use constructive logic, several use straight ZFC/HOL. Amongst those that use constructive logic it is usually only the kernel (containing the Type System) which models it while the rest of the system still uses classical logic.
They prove its consistency relative to another mathematical theory we assume to be consistent. That is if that mathematical theory is consistent then so is the kernel. In the case of Lean's con-leche and con-ron, it is proved consistent relative to ZFC + infinitely many inaccessible cardinals.
ZFC + infinitely many inaccessible cardinals is a fairly weak extension of ZFC widely thought by set theorists to be consistent. This is especially the case if ZFC is consistent (although by Godel we can't know for sure!).
ZFC + infinitely many inaccessible cardinals being inconsistent would be a really big deal! Given how confident set theorists are about inaccessible cardinals most likely it would mean that ZFC itself is inconsistent.
I feel like the real answer has gotten buried here, so I will answer my own question. The kernels are not "proven consistent" in any absolute sense. They are proven equiconsistent with ZFC plus inaccessible cardinals. This means it is "as consistent as everything else we have been doing for the last century," which is good enough.
Thinking more - as more such redundant tools are developed, surely someone will run all the existing big proofs against them, including this one. So we'll have more confirmations over time.
> the proof exploiting a bug is vanishingly small now.
Why do you say this? LLMs have been very recently finding all sorts of heretofore unknown bugs and exploits for languages and applications that have been around for decades.
There’s being careful and there’s being conspiratorial. At this point, believing that there’s multiple soundness errors in every Lean kernel that could all be exploited simultaneously leans heavily toward the latter.
To be clear, that doesn’t mean it isn’t possible. It just means that it’s extremely unlikely and we have no a priori reason to to believe it’s plausible or even possible to do.
> While still possible, the proof exploiting a bug is vanishingly small now
That's not how mathematical proofs work, though. What you have proven is not RH but rather the statement "either Lean has a bug or RH is true". So at this point you might as well be celebrating a Lean fuzzer.
There are multiple independent implementations of the Lean kernel.
“Either RH is true or every separate Lean implementation—some of which have their own soundness proofs—has a soundness error that can be independently exploited without triggering errors in any of the other Lean implementations.”
For traditional proof, it means "either X is true, or the author and the editor and the peer reviewer missed the mistake". I've seen errors in math articles published in serious journals, mostly at the typo level.
My understanding is that the Lean kernels only demonstrate that the Lean expression is correct.
The actual Lean proof (written by human + AI) still goes through a compiler to generate the Lean expression. So it’s still possible there are compiler bugs.
I’m hoping someone with a deeper understanding of this system can dive in.
If I understand correctly, for the purpose of the proof Claude formalized a huge chunk of all modern math. So it's a big result not only for GRH, but math foundations and formalization.
Yeah I’m wondering how on Earth a single conjecture can require 17 million lines of code to prove, but if it involved building up an entire ecosystem of theorems that had not yet been formalized then I suppose it makes sense.
I don’t know about anyone else, but I’m yet to come back from a coffee break and find my agent has solved a millennium math problem. I guess there’s always P vs NP?
The proof has also been built in parallel, so it's possible to mechanically shrink it a lot. Some equivalent lemmas have been proven more than a hundred times independently because different agents needed them.
The main theorem's cluster directly imports only three others: a zero-free region, primes in progressions and Mertens-type zero sums.
Most of the code sits 3–5 import steps below those, and no theme is more than about 4% of the code.
(I made the map. Cluster names are written by Claude from code samples, so they can be wrong.)
this is cool! Not sure this is up your alley, but I built a tool so anyone can run the lean proof at home, and there's a global leaderboard like SETI@home
Come join! Maybe I can integrate some of your visualization into our map.
> ...if it involved building up an entire ecosystem of theorems that had not yet been formalized then I suppose it makes sense.
This is pretty much it. Pre LLMs very little maths had been formally verified and a lot of it relied on proofs from humans convincing other humans, and issues did arise which would occasionally mean chunks of maths we previously thought were true had to be reconsidered.
Proving something in lean probably involves writing a lot of proofs for new areas, and the formal proofs are usually significantly more verbose than a human proof where a lot is often left for the reader.
Lawmakers often make laws and regulations that aren't consistent with existing body of law, but it's sometimes hard to see (esp. if they're doing a tight schedule, like having a 500-page document dropped in their lap over the weekend).
Having a way to say "here, this new bill you're proposing is provably in conflict with those three" would maaaybe nudge some people to take a second look or raise concerns?
It would also be nice at the judiciary appelate level if you could ask "does this particular finding conflict with law" directly as opposed to "do you, judge, as a fallable human being, agree with this/these other judge/s" in a turtles all the way down to the supreme court which is just a group of fallable humans rather than an individual.
Quickly knowing which existing laws you're treading on while writing a new one, and then mapping them to precedents... that sounds like it might be useful feedback for somebody embarking on a legal refactor.
This all misses the point. Lawyers have jobs exactly because laws are intentionally written to be vague and have loopholes, even contradictions. That field wants human in the loop.
Sure, but if you're drafting legislation you want to write with an understanding of how the human in the loop is likely to behave. What moves are going to be within easy reach? What other moves are going to be a stretch? They're going to do the research to discover their options, why make that research any more difficult than it needs to be? Why not associate it with a hotkey?
The outcome will be the same in either case, so you might as well make it quicker and easier.
Unfortunately, principle of explosion means that if the legal system has a single contradiction, then every law both contradicts and agrees with every other law.
"Yes, you avoided running over a family, but you crossed over the yellow line, didn't signal properly, that also counts as reckless driving and a number of other violations. You'll be fined thousands of dollars, your license taken, and you'll spend one week in jail"
Why would it lead to that? There are many structural errors, inconsistent terms, unintended gaps. Fixing them doesn’t mean that judges will be less flexible in circumstances where the law was supposed to rely on their judgment.
We are talking about people seriously arguing over how to interpret a comma, or the lack thereof [1], or claiming that obstructing guards outside a military base is a crime, while doing it inside the base is completely legal [2].
It's going to be fun when legal systems are analyzed by LLMs. The chances that hundreds of thousands (millions?) of laws and regulations written by legislators, judges and bureaucrats over hundreds of years are self consistent and contain no internal contradictions is very, very, very close to zero.
Laws aren't supposed to be strict logical statements that contain no internal contradictions. Laws are intended to be interpreted by judges for a reason.
It benefits the police to have a population composed of people who are breaking the law, even unknowingly, since that means any inhumane action by police is still "fighting criminals".
On the other hand police unions have obtained "qualified immunity" which means ignorance of the law is a defense for police.
Which is itself a demonstration of how contradictions can be beneficial.
Imagine the volume of exploits we have in software, but in the law. It’s happening in the law world, just slower and we barely hear about it here on HN.
It’s going to be a proper mess before it gets better.
Laws are supposed to be applied (by judges, bureaucrats, etc) with discretion and common sense, not mechanically. Of course, that can sometimes result in injustices – exceptions being made for some people not others for reasons of political or societal biases – but it also avoids a lot of injustices: thoughtless/mechanical application of the law can result in outcomes which the legislator never intended and few would consider just or fair or reasonable.
I think, if we actually started applying laws in such a mechanical way, we'd end up with laws changing a lot more frequently, and with a lot more special cases written in to them – now, if applying a law in a certain situation seems unreasonable, judges will often choose not to apply it; with mechanical application, they'd have no choice, and then the people negatively impacted would lobby the legislature for a change in the legal text, and they'd probably get it. It could also see increased use of executive clemency (the power of pardon), or its legislative equivalent (legislation retroactively declaring certain convictions null and void).
Again, I think you are missing the point that many instances the ambiguity in law is entirely intentional and by design, either with the intent to defer these ambiguous cases to the discretion of judges / admin agencies or as a part of political compromises or because those exact ambiguous/edge cases can't be agreed upon, or with the intent of future proofing. The specific example you gave is potentially exemplary. Don't assume that there aren't any lawmakers that have not taken a formal logic course
So the clarity you are proposing would in many instances have rather brutal and likely unintended impacts if broadly applied.
Not to mention that rewriting that verbiage probably wouldn't even have the intended effect as the semantics may already be determined by years of caselaw. But regardless, the dominating factor here is that the ambiguity is (usually or at least very often) baked in and that is in part meant as a stopgap against the particular kinds of injustices that arise when laws are applied purely procedurally without regard for the situational reality
Ambiguous "and" versus "or" is not future proofing and causes nothing but pain.
If they actually couldn't decide on such a thing, they would need to suck it up and do their job. Your idea of brutal consequences is ridiculous in a situation like that.
And as I said before you can still have ambiguity and discretion with a good formal wording. But it's clearer where it is.
> Ambiguous "and" versus "or" is not future proofing and causes nothing but pain.
We don't need formalising law in something like Lean to get this.
All we need is better legislative drafting. Indeed, many jurisdictions now have drafting guidelines which recommend using constructs such as "At least one of the following", "All of the following", etc, to remove the ambiguity of plain "and"/"or". Those jurisdictions which don't have such guidance should adopt it.
"law can result in outcomes which the legislator never intended"
But what has always baffled me, AFAIK, the original legislators don't file Amicus Briefs when the meaning of laws are argued in court. I mean in situations where the original legislator is still alive and of good mind, why doesn't the Judge just ask them what the spirt of the law is?
The parliament as a whole holds legislative powers, not just the person who wrote the proposal for the act. That's the actual interpretation that matters.
I believe they were talking about humans correcting human proofs pre-LLMs enabling us to define things in Lean at a useful scale.
But, to your question, I don't think any important error findings due to Lean have been unrepairable yet (for some definition of important). So far the ones which have really failed to be patched or been plain incorrect conclusions have, thankfully, had a very small blast radius.
The four colour theorem is a nice example. There was an early proof from Alfred Kempe which stood for 10 years until a flaw was found. [1] It was later proved using a different technique, so the status of the theorem went roughly
Yes, it's a nice small example to illustrate the point. Perhaps "chunks" was overstating it a bit, fortunately it's unusual to build lots of new maths on an unstable foundation, but finding problems in proofs is somewhat common.
it doesnt. a cursory pass suggests 3-9 million of these LOC are duplicate or dead on the low end. it'll probably be faster to remove the dead code than to run it end to end in its current state against everything.
I didn't check but I'm a bit skeptical, I'd struggle to believe that they wouldn't have done at least a surface clean-up of the code before publishing this massive breakthrough. Dead code is the lowest-hanging of fruits, hard to believe they wouldn't have bothered to prompt an agent "remove dead code".
i dont think dead/unused/duplicated code matters as much as "does it lean?" so waiting around to pretty up the codebase would probably draw some ire. i mean, whether its 7 million lines or 17 million, who's really gonna read it?
Unfortunately, it is not. The same level of slop nuke occurred with Anthropic's auto-formalization of Fermat's Last Theorem (~13 million lines of code). What Anthropic's FLT did do was undercut the human-led efforts to get idiomatic formalization of the requisite mathematics surrounding Fermat's Last Theorem into Mathlib.
This result is not quite as harmful as Anthropic's FLT was because it is indeed telling us something new -- that the Riemann Hypothesis is true, with full confidence -- but it is not reusable in any sort of way, beyond proof mining to figure out what in the world the AI model did. It is still in fact harmful to the greater project of mathematics formalization (irrelevant to foundations), because it undercuts the ability to fund such efforts for reusable proofs.
One indeed might wonder whether the "human-led" efforts of Mathlib are worthwhile. First, it must be noted that Mathlib serves two purposes -- to be a reusable and robust collection of formally-verified mathematical statements, but also to be a learning resource for students of mathematics without access to the rigour that an academic education provides. It is for this reason that the human nature of mathematical exposition is deeply important. In mathematics, a proof is more than a blackbox -- it is (or was) the way in which you explain mathematics.
(I consider "proof" somewhat unfortunate terminology for these slop nukes, as such. I think the term "certificate" is better. 17 million lines of Lean does not pragmatically convey anything other than whether a statement is true or false. It may be easier communication, but it is deeply, deeply, inefficient communication.)
The first goal of Mathlib is indeed not in conflict with slop nukes, however. One of the projects of the Lean FRO (or just a Lean FRO member? unclear to me how endorsed it is) is a version of Mathlib that fully embraces slop: https://taucetiproject.org/. However, there's really no bottleneck there. If there is any non-novel mathematics in Anthropic's proof of the Riemann Hypothesis that is not already in Tau Ceti, it is simply because they have not bothered yet. Auto-formalization works, at scale, very well, as harmful to the human as it might be.
Yes, you can. I know of a number of mathematics PhD students who got scooped by OpenAI's recent slop nuke. As such, they dropped out -- that the problems they were working towards suddenly became resolved meant that their dissertation topic suddenly was no longer sufficient, and their past work towards it of far less academic value. It is certainly one's choice to stop work on something, but this "choice" is intensely tied to their material conditions.
Aside from dissertation material no longer being suitable, there are external factors, being the ability to get grants. It is much harder to get a grant to understand and produce exposition of an already-resolved problem than it is to get a grant for a program towards resolving an open problem.
Anthropic is not the only company that has been harmful here. Math, Inc. kicked it off way back in March by scooping Jeremy Avigad's student on the mechanization of Maryna Viazovska's sphere-packing results. You can read Avigad's thoughts here: https://arxiv.org/html/2603.03684v3. It is obviously deeply harmful to his student, who suddenly found himself without a masters' thesis. What I find deeply disappointing is that it has been clear auto-formalization is a cancer since the Math, Inc. fiasco, and yet... the AI companies have persisted. Anthropic knew very well that what they were doing with FLT was undercutting Kevin Buzzard's program and everyone involved in it, and yet they decided to slop it out regardless, in pure pursuit of capital. Evil. May these companies all crash and burn.
Yes indeed I am arguing semantics, because I think it’s bad to define “harm” this way. For instance, if this definition were applied in tort law, you could sue someone for selling a competing product and reducing your sales.
It's so crazy, seeing all these claims that the only way for us to fulfill our potential as humans is to do a robot's job.
If I were one of those math students "scooped" by a "slop nuke" as you put it, my reaction would be gratitude for saving me from wasting any more time on work that isn't necessary. I could then turn my attention towards more interesting problems.
Unsurprisingly, like any good mathematical or scientific work, this proof has left us with more problems to tackle, not fewer.
Yeah, I'm sure you'd be sooooo happy to be a fifth-year PhD student, and suddenly have your dissertation work done (in passing) by OpenAI. Practically ecstatic, I'm sure.
Indeed. I was referring to the OpenAI "mathpocalypse" (https://en.wikipedia.org/wiki/OpenAI_bulk_release_of_mathema...), to be clear, although I find it not unlikely that the Anthropic proof will be found to have resolved various smaller open problems that will snipe yet more mathematicians out of a dissertation.
The OpenAI announcement was the particularly devastating one for the mathematics community, particularly in its clear statement to continue to milk the mathematics community for open problems to use as benchmarks. There are some deeply emotional remarks from mathematicians on this topic here: https://proofsandprompts.com/2026/10/08/100-reactions-to-100...
Hmm, proofsandprompts is indeed a good place to spend an hour or so. Thanks for the pointer to that.
I was vaguely aware that OpenAI had dropped some papers last month, but I didn't know it was such a big deal in the mathematics community. One can only imagine how the people who contributed to that page are reacting today.
It feels as if knowledge domains like Math will be forced to become closed-source with Oaths and NDA's and security scanners (no AI agents/terminals on your device), preventing open sharing of knowledge. AI company wants access to Math proofs ? Sure: $100,000 per paper, pretty please.
Not sure what you mean by "gatekeep knowledge" -- in reference to Alessandro Masullo's proof, I assume?
The original proof of Fermat's Last Theorem was ~200 pages and took experts 2.5 years to review. The (possibly correct) human proof by Masullo is ~1000 pages. It will take years to review its prose exposition, regardless of a corresponding Lean mechanization being developed. The one year they intended to keep it under wraps is short as far as mathematical timelines go -- and solely for the purpose of not hyping up a proof before how likely it was to be correct was assessed. But really, to be quite frank, mathematicians aren't "gatekeeping knowledge" from anybody here. You will never understand any proof of the Riemann hypothesis. (Neither will I. Neither will anyone who is not an expert in analytic number theory.)
As for stifling progress, perhaps -- but progress for what sake?
Paraphrasing, we're told that the reason mathematics gets funding is that the results can be used in other fields for problems we haven't encountered yet. If the results are only useful for expert navel gazing, then 1) let's revisit the funding issue and 2) have you actually lost anything? Fewer people able to afford to live producing proofs - for themselves?
Exactly. I'm not a research mathematician, but I studied math in university, and I have a hobby-level interest in pure math. So it's safe to say I'm more interested in math than 99% of the population. If _I_ find these "hooman math" arguments unconvincing, then chances are that 99% of population would be like "you want to get funding for WHAT?".
E.g. Hardy compares pure math with poetry - but are poets publicly funded? Very few are. I don't think it would be in interest of hooman mathematicians to bring up these arguments in discussion with normies.
And a zillion page AI proof is far more likely to only be useful for navel gazing, while displacing the creation of the better version because the problem is "already solved".
In reference to your idea that the AI companies are doing something wrong here.
As for Masullo's attempted proof, that story fell apart rather quickly, did it not? Of course, I can't say with absolute certainty that it can't be salvaged, but I'm not holding my breath.
If what you are saying is true, then that would make me much less likely to support math research. Because the implication of what you’re saying is that the majority of the benefit society gets from proving these theorems is minting one new math major.
Now I’ve heard people criticize the AI-proven theorems as not being done in the right way, not inventing any new tools, or being slop themselves. If that’s the case, the theorem can still be proven the right way. Or the researcher can refactor the code formalizing the new tools. If this isn’t enough for a PhD then that is the fault of the math community.
Either way, if what you say is true, that AI proofs of theorems are on net harmful. Then there is something seriously wrong with mathematics.
> Because the implication of what you’re saying is that the majority of the benefit society gets from proving these theorems is minting one new math major.
These 'math majors' (actually doctoral and masters candidates in mathematics) work for the defense firms your government runs, the security apparatus of your state, finance firms -- anywhere there are a lot of complicated mathematics, that's where they are.
This is not like CS/SWE where four years of college makes someone competent for work. If we don't have these doctoral and masters students, there will be a vacuum in the next decade as the normal pace of hiring outstrips the number of qualified candidates.
It's not clear what will happen, maybe LLMs and other AI/ML tools will fill this vacuum. But this is about much more than 'math majors.'
Sorry I meant PhD and I specified PhD elsewhere in the comments.
If there really is demand, math departments will find a way to adapt. If there is seriously no way to train new mathematicians without open problems, then I don’t see how they would be equipped to add value in this world. But I don’t think that’s the case, they will figure it out.
I think you and your sibling commenter miss that doing these advanced jobs is not like flipping burgers, or even like doing financial research. There are problems that may not have fully-specified solutions, and the best way we know to train and identify those who can solve unspecified problems happens to be asking advanced-degree candidates to solve unspecified problems.
That is a reason why reasonable countries fund research, even when the research isn't going to yield obvious benefits on its own merit.
I think this the best triage of the problem's kernel. Mechanisation of math/any research is mechanisation of thought. So, unless we believe in times of crises, big or small, instead of summoning the obscure symbologist, we can spin up an agent instead, that will then have the requisite grasp for the problem domain to delineate the most fruitful lines of attack, (in which case, all humans are outmodded already) we would need the experimental archaeologists, xenobotanists, and garbologists.
I agree that it's heartbreaking, but any advisor worth his/her salt would find a twist to publish some of the side quest or a followup.
It may be harder to get your deram position immediately, but that also requires a few results and good recomendation letters from relevant people, not just a thesis.
If you spit on someone's plate and he stops eating, you didn't harm him, because it was his choice to stop eating?
Publishing a proof demotivates other people from working on their proof. It would be false dichotomy to say there is now no motivation to work on the proof, but the motivation did get smaller. How much smaller is hard to assess because it depends so much on human psychology and it's all new to us - if someone publishes an "idiomatic", easily understandable proof, will he be praised similarly as if AI didn't beat him to it, or will people handwave him as simply explaining what AI already figured out? Who knows, maybe he did very little work and just asked chatGPT to explain the existing proof? And even if that won't happen, people working on proofs might assume it will happen and therefore move on to other things.
Not sure if you're mocking the other person that replied to my comment, but if not: that's the thing with analogies, they are always a different situation. See my other reply.
The purpose of an analogy is not to present a completely equivalent situation (though that's best), but to present the same logic in a different situation, so that those who failed to see the logic in the first one can hopefully spot it in the second one.
It's fine that you still don't see the logic, though your comment was completely unproductive.
Everything changed this fall and what is being funded must change alongside it. Explaining, simplifying and understanding these sloproofs may be very well more valuable than trying to prove the conjectures.
I agree with this, to be clear! It is deeply sad, but I really don't think mathematical "taste" is an inherently human notion. There are plenty of raw metrics of quality to optimize proof certificates against.
I am making the specific statement that Anthropic's proof of the Riemann hypothesis here is not of use for "math foundations and formalization", because the auto-formalization process of existing mathematical results is no longer hard, as shown by projects like Tau Ceti. What the greater project of mathematical formalization needs is proofs of high quality, suitable for inclusion in Mathlib. There's a reason Tau Ceti exists alongside Mathlib, despite Mathlib accepting AI-generated submissions. Cleaning up 17 millions of lines of garbage of this Riemann hypothesis proof is simply not worth it re: formalization, because formalizations of existing mathematics can just be... generated.
And yes, this is novel mathematics. The proof strategies will likely be mined and lead to insights and new proof strategies that human mathematicians can use, too. It is better than Anthropic's FLT formalization, which was entirely worthless. But that's not to say that this isn't a sad and disgustingly exposited result, that was produced as propaganda for Anthropic, and (considering the lines of human research cut off by having a result) a net harm for the mathematical community.
I am not in the front row, but what you say alligns with how I would assume things play out, unfortunately (and short-sightedly; how easy would it have been to not antagonize the mathematical community, i.e. by redirecting a few 10M$ into scholarships, conferences etc... also maybe 1-2 fulltime positions that focus on human-LLM-collaboration as in "how can we this new technology to increase understanding?", which seems to still be an afterthought for the big AI companies...)
Slightly related: I am fairly confident that we can let LLMs loose on those 17M lines of code with the instruction to refactor the heck out of it: Instead of "optimize that program so that it runs faster" (which AlphaCode did to great success already), the task would be "optimize that program so that it uses less theorems" etc... which is a very easy-to-verify task, since we after each refactor, the end result will never change! I predict that 1) this will be done more; once this first wave settled down a bit 2) this too might be a way to help us (humans) discover and understand new mathematics. For the LLM, it might have been just a boring ordinary code refactoring by extracting some common code out; for mathematitians, this might actually correspond to a deep connection between two areas.
Which is pretty cool IMO, it's basically an army of statistical parots that mine abstractions all day, and they only can succeed because at the end it is still the same outer lean statement being checked (we just rearange some stuff under the hood)
I am 100% sure that yes, it's just not mainstream yet. (Why? Because lean proofs are computer programs)
----------------
The big caveat: This is only if the proof is formulated as a computer program, which is the case here with the Riemann Hypothesis proof. (Lean is a general programming language; and every "proof formalized in Lean" is also a Lean computer program.)
Much of mathematics is not yet formualated in such way (although much progress was made by countless contributions of researchers and amateurs in the last years, both in Lean but also in other proof assistants. Lean seems to become the de facto standard though, but I am watching from the sidelines, so take that with a pinch of salt please)
Yes, absolutely. You might take part of your big proof and extract it out in a named lemma. Then you might find that if you strengthen this lemma a bit, you can suddenly use it a multiple places in your main proof (the analogy here to programming and extracting things into generalised functions should be obvious). But you might also take whole subparts of your proof and prove it in a different ways, but as long as the assumptions and conclusions (input / output) stays compatible, it does not need to change much in the main proof.
Won't it be possible to de-slopify it incrementally?
E.g. as long as you have compatible definitions, you can literally go theorem by theorem. That should be easier than refactoring software because you don't need agreement on things like state management and concurrency.
Regarding funding, I think quite likely AI labs would be interested to pay for better proof formalization as that can help to train models to make less slop. Isn't that better than make taxpayers fund this work?
What's the state of proving for these proofs? As I was working on Riemann@home so people could run the proof on their home computer, I was checking out options involving RISC Zero and zkVMs. It looked like it would take about 26 years of CPU time to prove this Riemann proof in a zkVM. Which seems like a lot to me. But once verified it can be confirmed in a second.
What would need to be true for this type of computation level to be valuable to the field?
I don't think it's valuable for math, but for software - maybe. E.g. fully audited supply chain. Install new OS if there's a proof it doesn't introduce new vulns, or something.
Interesting! And with math increasingly becoming software, I wonder how many more proofs like this I’ll find interesting enough to run on my own computer.
A fake refutation of the Collatz conjecture already got through one of these independent checkers in the past.
What's harder, proving GRH or finding a way to get through every checker (which aren't as independent as they are sometimes touted to be)?
In this case there was a human expert in the loop so it's much less likely the model hacked its way through the problem without the expert noticing. But the statement as written has 4 points, none of which rests on the human involvement. So I don't like that they claim full confidence based on those points.
In short, Lean has formally verified checkers now, and the RH proof was checked with them. So any such soundness hole would have to be exploited in the correctness proofs of the checkers, which is a much smaller attack surface than the RH formalization.
One thing that every manual/automated verification relies on is the consistency of ZFC + few inaccessible cardinals. I mean while humans have carefully thought about the flaws in that for century and didn't find any, it's not grounded in any clear reasoning why it is consistent.
Not at all saying it is inconsistent, but I personally find the liberal use of introducing inaccessible cardinals axioms a bit concerning.
Much of mathematics relies on these axioms. There is nothing special about the dependency of proof checkers or verification of this particular hypothesis. What is the purpose of this statement in this context?
Many set theories were proven to inconsistent, and lot more type theories(e.g. ZFC plus a Reinhardt cardinal, Martin-Löf's first type theory). And if in AI posttraining it hit it once, the probability of hitting it increases because of reward hacking.
It's possible that the question of consistency is too abstract to judge its significance. Why?
(1) A lot of the mathematics that was done in the 18th century was
obviously inconsistent and yet still useful.
(2) Trying to verify logical consistency of a theory is basically impossible
because of Goedel's 2nd Incompleteness Theorem.
Hitting a wall like this makes me think we're asking the wrong question.
Most natural problems like RH, twin primes, etc dont fundamentally rely on the inaccessible cardinals. Any problem that depends on them, there is much debate whether they can be called natural.
Many problems may not require even full zfc. zf, peano or even weaker may be enough.
There are also many theorem provers not just lean. lean depends on type theory, but metamath for example depends on zfc directly.
It may be a good but likely tedious exercise to convince yourself by translating the current proof to a zfc dependent prover.
Yes it does. Sets could be represented in Martin Lof type theory and all the ZFC axioms proven inside it. If ZFC is inconsistent, then so would be the type theory.
> I mean while humans have carefully thought about the flaws in that for century and didn't find any, it's not grounded in any clear reasoning why it is consistent.
The reason on how they are stressing, that the original definition of the Riemann Hypothesis they are starting from, is very sound ... tells you everything you need to know...
It's turtles all the way down (what if every x86 and ARM chip has some obscure ISA exploit that agents are silently using that we don't know about?) but the idea is that while you can't remove absolute dependence, you can remove chains of dependencies and keep making it smaller and smaller w/ reductionism
Yeah, I'd just add here that the top-level con-leche is proven correct with respect to a set-theoretic model of Lean. That model very well could be inconsistent! But it's unlikely. What con-leche et al. do is eliminate trust at other levels. con-leche removes the need to trust Lean's typechecker, con-ron lets you trust either Lean's compiler or Rust's compiler, con-agua lets you trust only the Kraken x86 semantics. Turtles indeed. No way around it!
The challenge would be encoding the ISA exploit into the Lean proof in such a way that the CPU would be able to recognize it and behave in an exploitable way, without also making this complex behavior stand out in the (surely short and readable :-) proof text.
Then someone tries to verify the proof on some other kind of CPU and the differential execution trace shines a giant spotlight right at your ultra-valuable backdoor.
But extremely risky too. The only thing needed for someone to expose your backdoor is for them to run the lean compilation on an older CPU (or different CPU) without the backdoor, and suddenly there are very knowledgeable people interested in finding the embarrassing reason for the observed discrepancy.
You can keep writing simpler checkers that check the more complicated checkers. But at some point you will have to give up and say "this checker looks correctly implemented (to us) and gives the results we expect on all our tests."
Depending on how the checker works, you might be able to run it on itself. But that doesn't say much; maybe it only approves itself due to a bug.
> But the statement as written has 4 points, none of which rests on the human involvement.
The theory and implementation of Lean is very much so a result of human involvement, on a massive scale even. Just because it is something repeatable does not lower that in any way.
Distrust of LLM output should not turn into distrust of automation, I think. The former I totally get (and deeply hold myself), the later is likely to hold one back - imagine banning calculators because there is not human in the loop, to make an extreme example.
I'm not throwing any doubt on the reported results, your comment just provokes a further question: How do you know what the LLM "intends" and whether it might/might not ultimately be interpreted by us as malicious?
We need to give Levent a lot of credit for this. As much as we talk about Claude proving it, Levent seems to have a rare talent for getting the most out of it.
Should we though? Is working at anthropic and using unrestricted unreleased models really fair?
If everyone got to do that then we would really know how rare his talent is. But we don't, and he doesn't post his prompts or workflows. So how can you be so sure.
If he was at least posting his prompts we could acknowledge that. But honestly I'm not going to give him the benefit of the doubt on this. Highly accomplished and successful yes, but uniquely talent who knows.
> Should we though? Is working at anthropic and using unrestricted unreleased models really fair?
Most of the discoveries in the last 2000 years were made by individuals who were born into wealth, had the right set of circumstances, or were blessed by the powers that be at the time.
Imagine all the smart minds throughout human history who toiled away in fields collecting food and will never be known about.
Unfortunately scientific discovery has always been unevenly distributed.
"I am, somehow, less interested in the weight and convolutions of Einstein's brain than in the near certainty that people of equal talent have lived and died in cotton fields and sweatshops." - Stephen Jay Gould
Is Alpöge from some extraordinary rich background? Otherwise I don't know how this is relevant to the topic at hand. He got a job at Anthropic, likely based on merit, unless someone can state something contradicting that.
Your argument seems to turn the discussion ad hominem. Social affects could be discussed in a way that does not do that, as in: The style of discrediting critique by invoking the antisocial affect of feeling superior/trying to make the other feel inferior is the kind of antisocial affect the forces that tear up the social fabric thrive on. I'd argue your comment does not and contents itself with being a feedstyle flame.
But that’s not what the discussion was about, that’s irrelevant. It wasn’t about whether Alpöge was rich. It was about whether extraordinary resources are required in this age of maths. There’s a subtle difference, and certainly we can talk about an overlap, since wealthy individuals generally have greater direct access to extraordinary resources. But a single individual’s provenance doesn’t constitute a strong argument against the general point.
It’s a bit disingenuous to seize on the equivocation caused by the topic drifting like that.
We'll never know, you state it with certainty but you have no way of knowing either.
You're probably just lying to yourself for the comfort of believing the world is "fair." It's not. It's extremely unfair, and most of the time it's just average individuals with access to unfair resources that allow them to exploit everyone else.
This is obviously apples and oranges - Levent will get the credit but with the asterisks that he used AI and these people will not have that asterisks. Seems fair to me.
An equally interesting question would be, how many Ramanujans do we need?
To what extent they could be actively contributing to orthogonal math directions isn't clear to me. To some extent, new findings are constrained by time itself passing, and new generations being taught on pre-existing knowledge.
And in 21st century, we could say we’ll finally bring all the smart minds out of poverty and use their potential to progress the human civilization, but lucky us, creating clankers to replace them was cheaper and more convenient.
Or rather, for $20/month some savant somewhere will get a partner to bounce their creativity off of whilst living in their lower socioeconomic system.
For example, my single mother mom cleaned houses and worked as a secretary. She never graduated high school and had no interests outside of television. At 9 years old I asked for a C++ programming book and started making games on the library computer. If I had Claude I would’ve save myself a lot of time asking Jeeves or random programming IRC channels.
Today, I used Claude to discuss the physics within black holes, theorizing about the collisions between two black holes, etc. It serves as a muse and a thought partner.
Anything but actually fixing inequality and improving people’s lives!
“We decided that actually improving things wasn’t economically beneficial for us, but if you feel like paying money to your corporate overlords you can tinker with their toys so that can achieve something which will never materially benefit you or your community!”
> If I had Claude I would’ve save myself a lot of time asking Jeeves or random programming IRC channels.
Isn't this what all of us wanted? A personal all-knowing Librarian! But we assumed only we would have access to it, and we would do extraordinary things, only to realize that when it comes, the rest of the world would change alongside it. That there would no longer be any need to do what I expected to do so well with it.
That resonates with me a lot. lately I find myself worrying that the librarian is turning into some sort of legislative entity. I know all the books, but I wont tell you where that knowledge came from. And this I will not answer for safety reasons etc.
The social restrictions, and access to knowledge are reimplemented right there.
Might just be that we wittness the reemergence of a priest cast in the sensu stricto, when that all-knowing librarian will be aligned with capital autopoiesis.
Those individuals over the 2000 years, however unjustly privileged, relied on their biological minds to achieve extraordinary insights. Much of Levent's cognitive effort here was partitioned with an automated language model.
>relied on their biological minds to achieve extraordinary insights.
No they did not. I'm not going to speak for Maths, but this just isn't true for physics. A huge amount of phsyics rests on the invention of the telescope, which required glass. Intuiting that light travels was incredibly difficult without the studying the motion of planets.
We're talking about automated cognition versus biological cognition. I'm not sure how you think a telescope performs complex cognitive labour? It is merely an instrument, where a language model is a highly performant second brain (and please don't reply with the "but muh abacus" false equivalence).
With a finding like this proof, nobody is saying you aren't brilliant. But you are essentially setting a 5 mile track record by completing part of your run on a moped.
Yes, so the computer (and all the engineers that created it, which includes Levant to some extent) gets >99% of the credit and Levant gets <1% as the prompter.
This is true, and unfair, and can make one angry, but its also kind of inescapable. Imagine a system where anyone could easily do such things, starting a large business or scientific lab, or proving the RH, for example. In such a system, if any person can do those things, they have little value / are not noteworthy. If someone has access to more resources, training, etc, then they will also be able to do those things, but bring more resources to bear (people, money, time), and therefore their creation will have more "value". As a result, whatever accrues to these endeavors as a result of "value" will go more towards the person with those extra resources, and cause a rich-get-richer spiral.
Its kind of an inescapable feature of free markets and trade, regardless of how high the floor gets, differentiated inputs create differentiated outputs. Even if the input is "one person has a far superior mind", it might create a difference, even if everyone is free to do something.
The HN crowd seems to be biased towards thinking that someone should receive outsized rewards due to their inherent intelligence (a matter of luck and genetics), but not other lucky factors such as looks, connections, family name, inheritance, timing, etc.
Is it lucky that these people had access to unlimited compute and cutting edge models no one else did, and time/resources to work on Riemann? Yes. Does that mean probably tens of thousands / millions of people, if given the same chance, could do it? Yes. Is that fair? No. Is there a way around that? No (for the reasons I gave above).
The counter-argument would be "everyone should have access to models with all these powers", in which case "everyone can solve the RH", in which case "solving the RH isn't impressive / hard", and only solving something most people can't solve is noteworthy, and hence anyone with even more compute would be seen as "winning unfairly".
To take a more neutral example: music. Most anyone can make amazing music today with VSTs and DAWs. So baseline music has little value (whereas hundreds of years ago just hearing/making quality instruments was a rarity). But some people bring more musical skill due to training, background, genetics, exposure, mentorship, practice, etc. All of those things require resources, luck, and circumstances. They also need time and resources to work on their compositions. Some, like Hans Zimmer, were lucky enough to land on some big movies. Are there thousands of musicians equal to or more talented than Zimmer? Yes. But like I stated above, its an escapable feature of a system where something like "can bring more value because got lucky movie deal" is a thing, if "can make better music because talent" is a thing.
I know I gave a long rambling response, but I think dwelling on how the successful were already winning can be depressing. It can be helpful to become more stoic about it by realizing its kind of just the way things are (resignation and acceptance). Do I like realizing that there is probably nothing I can do to create the next FAANG? No. But its also freeing / helpful to realize that there is a range of what one can do in life based on the system, and to thrive / succeed with the range one is given, taking what opportunities are available.
People are sad the RH was solved because it was a machine / privileged person. This is how humanity works, countless human lives filter through a giant genetic optimization algorithm, and a select few randomly hit on meaningful progress for the species. So hopefully we can just accept the blessing that progress / technology is being built by someone, and everyone (mostly) gets to have some of the fruit falling from the canopy.
This actually ties into a thought I was having yesterday about requiring the full prompt and response flow to be included as part of an academic submission that relied this heavily on AI to complete.
It's not just the proof that's valuable to scientific discovery, it's the process of using a powerful tool at an advanced level that others could learn from. It would also help to show concrete evidence of any sudden shifts in the AI that might indicate training data was siphoned from other researchers (which I don't necessarily think happened here, but is always an ongoing question).
Who we credit when things get done has to do with being fair with the work they and other people involved accomplished to get that thing done.
If you're not fair, people won't work. It's a problem for mathematicians and many other jobs. I don't think it's happening here, but you can do 95% of the job, and someone else with more money takes it and asks LLMs to do the remaining 5%.
In this case, on one side obviously levent alpoge is highly qualified, but the average mathematician also doesn't have $10m of claude credit to try to have the same result. So is it just about the money, or about the skills, or as always, both?
I think what's also frustrating is that LLMs look a bit like a smart bruteforce method. They can fail a million times and then get the right result because it's the only remaining solution. It's smart because the true number of things to try is infinitely bigger, but compared to how humans work it can feel like cheating a bit. Maths aren't just about bruteforcing solutions to problems, we also need to understand these solutions. And here I'm not sure anyone is able to understand it.
So, do we reward and applaude people solving things with AI if they can't understand it? If it's cancer or climate change I would still applaude, because I care about the result more than the "how" (which I do still care about). For the Riemann Hypothesis, I'm not sure. I care about the "how" more than the result.
To that point, it seems that he enjoyed interacting with Claude. Sometimes I wonder how many of the bad experiences result from what Jung would have called projections of the user's self onto the AI through their prompts.
"Skill issue" is very correct. For friends and people I like, I try to help them use AI more effectively. For mean and bitter people on the internet, let them be ineffective, I don't care.
I don't take those statements at face value - he's employed by them for a lot of money and is being given access to high level (and expensive) tools. What he put out was effectively a press release, and I don't doubt that a bunch of marketing, PR, and lawyers reviewed it before it was allowed to be posted.
Yes he is making money. Yes this was good publicity for Anthropic.
But this was not a press release. And I doubt that it got legal, marketing, or PR review before it was published. Maybe in the timing, but I see no internal evidence to support the theory that it was edited for marketing.
Nobody expects you to take it at face value, but you're erring in the opposite direction by dismissing it out of hand for these reasons.
As we say in Silicon Valley, no conflict, no interest.
There were certainly others at OpenAI and Google that were actively searching for a proof with frontier models. That he got there first is a credit to the intuition he had built up from decades of work on related proofs and deep understanding of related mathematical toolsets known to be relevant to different parts of it. That much is clear from his 4 page remark.
Unless there's reason to believe he generally lies, and given the fact that he's the only one to accomplish this so far, the default assumption is that it's not as simple as you're saying.
What? He just works at Anthropic, lol. Any monkey with a PhD in mathematics can do what he is doing given access to Anthropic's internal models... and they'd probably be doing a better job at exposition, too. Alpöge seems to be uniquely uninterested in explaining his results, for someone who calls themselves a mathematician. His approach has been pretty clearly harmful for the social structure of mathematics.
> Any monkey with a PhD in mathematics can do what he is doing given access to Anthropic's internal models
Clearly not true given that multiple labs have many PhDs attempting this and one guy got the Riemann Hypothesis, a large contribution to Navier-Stokes, and multiple others just in the past few months.
Yes it is. He is lying when he claims otherwise. He is a former professional mathematician who has announced all of Anthropic's serious mathematical results. He most certainly spends all his days pulling the lever on the internal slot machine.
I mean, do we? For all we know all he has done is had access to unlimited compute on higher power models than available to the general public. His general complete unwillingness to divulge any information about any of the results he has obtained (AFAIK, we are still waiting on any details for the Jacobian conjecture counterexample or high rank elliptic curves or Hadamard matrices or S^6 complex structure or...) combined with the general ability of current models to make progress with minimal prompting means the burden of proof is on him.
If he truly did have non-trivial mathematical input to this it should be trivial for him to give talks on at least the overall strategy of the proof. Before LLMs this was a bare minimum expectation for basically all mathematicians and their results, and I see no reason why we should give up on it now, especially so since the previous primary artifacts (the papers themselves) are now on their way to being completely automatable.
I mean, your average guy absolutely has no chance of producing anything of value without having a lot of natural intelligence, no matter how good LLMs are.
>Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.
>2 A prompt including similar encouragement was used to help Claude disprove the Jacobian conjecture.
Yes. This is what bothers me most about these "AI solved the Riemann hypothesis" headlines. Alpöge deserves the credit here. Mathematicians are using AI, just like they have been using computer simulations, calculators, and before that abacuses and compasses.
Totally true, I don't see why people suddenly downgrade it to solely Claude and start to criticize the entire things. Levent is the key with right intuition, background, resources.
Did we just witness the invention of the “telescope for mathematics”, starting with the OpenAI proofs and cemented in place by Alpöge? Is this the start of a new golden age of
Mathematics where researchers can peek deep into the logical universe, looking at distant proofs and infer other knownledge & ideas from it? It truely is an amazing age to be (come) a mathematician.
No, unfortunately not. The proof is filled with neuroslope and crazy ideas that somehow work.
Instead of simple and elegant theories, new geometry, or a unique approach, we've ended up with proofs that take up gigabytes of code and look more like randomly selected phrases assembled on a supercomputer and verified by a solver than anything useful. Wait a minute...
No lean is not good tools
look at mizar. this is complete and veryfy tools for mathematican, but it need loot of work, loot of puthing proof inside. https://mizar.uwb.edu.pl/
I don't think this is true, or perhaps the sentiment is far different on the applied math side. I am a computational mathematician and most of my colleagues are overall positive, I'd day.
Change is hard. No sense in fighting it, no sense in being bitter. Mathematicians will no doubt figure out what the future looks like for them just fine.
Do huge results like this open up even larger questions? Is there some question even more important that mathematicians dared not approach without doing big problems like Riemann first?
Always. They tend to do it by two mechanisms: creation of new techniques, and creation of linkages between techniques or disciplines.
the thing the current boom in AI mathematics presses on is the question of how to interpret the generated new techniques. So while we have a lot of new techniques put out there and a lot of new connections put out there, currently it's still mathematicians largely who will have to understand those and think about next steps and how to apply them. That won't be true for long, though, I bet.
But in any event the job of research mathematicians is changing really rapidly right now, and they are all sort of back to grad school in that there’s like 100 years of new math to catch up on in the last year.
I would add the caveat though that the newer questions sometimes end up becoming more and more abstract and less and less interesting sometimes. Like topos theory for example....
I often wonder whether the resistance many mathematicians show toward the use of AI stems not from the destruction it supposedly causes, but from the need for an entirely new way of working - new goals, a new environment, a new culture. One thing, at least, I do not see (yet?!): that the use of AI leaves mathematicians with less to do. My current impression is that they are needed more than ever.
That isn’t what they claim is being lost. Tao has explained this in detail: the problem is all their institutions are built around a particular method of generating signals of innovative and useful work that are now broken.
Yes, math will change and endure but it could be a small fraction of its current community size when it’s done. It could return to the era of dilettantes if no one can earn a living doing this new work.
The community could become very insular as the only defense against cranks.
Not sure there is any more a problem with cranks. It is easy to see now if what you do is correct, so the only question that remains, is it interesting? I see an explosion of math, especially applied math that is more pure than current pure math.
Have you been asleep? The problems of open source maintainers have made the front page dozens of times and the same problem is absolutely impacting Lean and MathLib.
The same story exists for Arxiv and that’s made the front page as well. Slop is everywhere. Just because Lean compiles doesn’t mean it contributes value.
You cannot compare open source software with this, because software has the correctness problem. Same for the arXiv, a PDF is not a mechanically checked artefact.
I don't think Lean needs to have a slop problem. If they do, they are managing it the wrong way. I think Tau Ceti [1] is pretty much their attempt at answering this problem.
He doesn't use that term or elaborate on why the destruction of the current mathematics ecosystem is a harm. He assumes the reader considers that an intrinsic harm. I'm describing what I think could emerge if no new sustainable path is found to train and employee mathematicians.
Another perspective shared by a decent number in the community (myself included) which is very low level and has no grand implications for the progress of math or society: doing math with LLMs is not fun. And many of us are incentivized to study this subject because it's - you guessed it! - fun.
I - and many other software developers can relate. The difference for me is I've always enjoyed both the craft aspect and the practical aspect of building useful products and tools, so there is some compensation.
I am absolutely not trying to diminish this aspect, more like just asking (as a non-mathematician): is it possible that simply not every problem is as fun fundamentally to solve? Like I would imagine a problem that has a million, trivial sub-cases is not necessarily as "fun" as one that requires some "creative insight".
From the fact that how large these semi-brute forced vibe proofs are, even assuming some very efficient optimization/shortcuts I would wager that it's still not gonna be an "eureka" proof we just somehow missed before.
As an analogue to programming, there are countless tasks that were historically just over the limit of automating, but were always boring "plumbing". I would wager most people never really enjoyed writing the boilerplate connecting the backend API to the frontend, and this was just offset with the joy of finishing a task, design decisions etc.
I think partially it’s because OpenAI (at least) was implicitly rather rude about it. It’s like building a new space telescope or particle accelerator and not putting a scientific committee in charge of how it’s used. Instead, throwing incomplete results over the wall for someone else to clean up. And furthermore making it clear that they don’t really care about winning prizes or getting papers published in prestigious journals.
So they’re beating mathematicians at their own game and implying by their actions that they don’t care all that much about the game. Rude!
And of course people try to rise above that and adapt but I imagine it still hurts.
But this latest release doesn’t seem rude because a prominent mathematician is heavily involved. (Though it inadvertently stepped on someone else’s toes since they had a proof in progress.)
Releasing just the Lean proof and collaborating on turning the raw data into something understandable seems like the way to go?
They have not changed their approach in any meaningful way. That they are continuing to use open mathematics problems to evaluate their internal models on -- something the mathematics community has directly asked them to stop doing -- means that AI companies continue to fundamentally undermine the mathematics community. There is no longer any motivation for PhD students to work on any particular open problems when they could be scooped at any time, and the AI companies being able to resolve open problems faster than mathematicians means there is no longer any particular ability to put programs towards resolving open problems in grant applications. It is a crisis of both funding and motivation that is being caused by the antisocial behaviour of these corporations. Their actions are not particularly defendable, in my opinion.
This just isn't true. What is true is that 40 years ago, most mathematicians used only pen and paper to do their work. When Mathematica was released in 1988 and started to gain adherents in the early 1990s, it was rapidly adopted by many as saving effort, but you could still decide if you wanted to only use pen and paper. I'd venture a guess and say as of five years ago very few people did not use some sort of symbolic math support in some way. Right now, nobody who wants to be a grad student in math can use pen and paper, except recreationally.
Mathematica as a tooling enhancement had a relatively slow impact, this tooling change has a big one. Grad students that don't want something to change rapidly are going to get f-ed hard right now. Nobody likes that. I bet you there's another group of grad students that are stealing Ritalin from all known sources and speed running everything they work on this month as well; from that perspective, the field just went wide, wide, wide open -- they will be able to put out more progress than their advisor's life work this month if they do a good job.
If I were a 22 y/o math grad student right now, I would be in heaven.
If you were a 22 y/o math grad student you'd be fearing for your livelihood and future right now -- the same as all of the SWEs that were "replaced."
Math has been blown wide open, but there's going to be a huge gap in enrollments + progress the same as we're seeing in computer science, until people adapt (3-5 years before the dusts settles).
The issue is with the second order effects and their consequences.
Math is finally embracing open source thinking and techniques for working at scale. It's been overdue and it will take the kind of collaboration physicists were forced into when they started to need big machines for their work. For what it's worth, formal verification probably means the open source model will work better for math than it ever did for software.
Mathematicians have already largely assumed RH/GRH is true. There's been debate over it, but it has been sort of 90% assumed true. All of those "if RH/GRH, then ..." results already build on top of it and were assumed to be on pretty good foundations, and now they're just all solid. It would have been much wilder if RH/GRH were disproven, which would have upset mathematics and invalidated all those claims.
I don't see how this result changes that much, since we woke up yesterday morning and the thing that most mathematicians assumed was true, turns out to be true.
And the 17 million line long Lean proof is almost useless in its current form. Until you refine that and get it down to a "proof that fits on a T-shirt" (or at least the length of a typical journal article), I think it is just screaming out that there are large missing areas of mathematics that we don't know about. It may be useful to "mine" the proof for theorems that could be (for example) used to add to the Langlands program, and as those get added to Lean, all that work could start to wipe out millions of those lines of Lean code. That might be done with AI/LLM-assitance, but it'll be motivated by human experts diving into the proofs, and none of it will attract any headlines.
I think the human obsession with "first post" claims on proving things we already expect to be true will need to get reined in. This isn't a P=NP kind of result that would blow up the world if we woke up some morning to find out that was true. And 99.9% of the really useful work on this proof is still in front of us.
And I suspect that using an LLM to find results like this is like nuking a fly in terms of the computational effort required. If we had better global optimization algorithms (genetic algorithms, particle swarm optimization, and techniques we don't know about yet) surrounding the problem space of Lean proving, we could probably do this with orders of magnitude less computational power. I doubt that the transformer neural network is the optimal architecture to solve these kinds of problems.
Thanks for articulating precisely what I've been feeling. I also wary of the arguments for whether this damages/progresses the broader human endeavor of mathematics--the answer is a little of both, and everyone involved would be better off by recognizing this ocean of gray, rather than drawing lines in their black/white sands.
> Do huge results like this open up even larger questions?
Yes. Questions like "when will the bubble pop and will it take down the US economy", and "is a cure for AI psychosis possible, or are we cursed with these psychos forever".
On the one hand, I really look forward to progress in physics to enable Star Trek-like technologies, but on the other hand, I wonder if we should solve our global conflicts first.
I mean, wouldn't it be great if AI could lead us to a world of lasting peace, prosperity, and happiness?
And go live where else? Mars is harsher than Antartica and warp drives depend on whether physics allows a feasible design that doesn't require too much energy. And even then, where is a habitable planet for humans out there?
> I mean, wouldn't it be great if AI could lead us to a world of lasting peace, prosperity, and happiness?
Not convinced that is within the capabilities of any AI without it doing things we would in fact not want it to do to us as a species.
We seem to be built for conflict and I'm not sure that what is left when that is removed is still us.
> Let me tell you something about Hew-mons, Nephew. They're a wonderful, friendly people, as long as their bellies are full and their holosuites are working. But take away their creature comforts, deprive them of food, sleep, sonic showers, put their lives in jeopardy over an extended period of time and those same friendly, intelligent, wonderful people... will become as nasty and as violent as the most bloodthirsty Klingon. You don't believe me? Look at those faces. Look in their eyes. (Quark from the always brilliant DS9).
OK, I'll bite: I find little support in history for the theory that conflict is inevitable because humans have an intrinsic desire or inherent nature that drives us to go to war. War is what people have turned to in roughly two circumstances, none of which are or will be ubiquitously present:
1. when (the leaders of some) some belligerent tribe, nation or coalition etc perceive that they can promote their interests by going to war without risks that at first blush will outweigh the gains.
2. when some (the leaders of some) tribe, nation or coalition etc can credibly anticipate that they are about to be attacked and then go to war, because 'attack is the best defense'.
Proof abounds in history that war is not inevitable. One example: peoples all over Europe that used to engage in constant war with their neighbors in the next valley, next district, next-door nation, over hunting-, grazing-, market- and raw material opportunities have ceased to do so even if they are essentially the same peoples.
It's of course essentially just simple cost/benefit analysis, not blind human nature.
I'm becoming increasingly convinced that many/most of the "great men of history" were actually insufferable ignorant pricks with astounding luck and zero concern for ethics or morals. This is mostly based on observations of the contemporary world, and how startlingly true it currently is. There's no reason to assume the people of the past, whose exploits were recorded into history, were remarkably better/different.
Below is a list of conflicts created, funded, armed, and backed by the USA. If you want to stop "global conflicts", you need to look in the mirror and understand who is the problem, before you propose any solutions, like stalling Physics research as a solution to global conflicts.
You need to have crystal clear clarity about who exactly is the Threat to the global peace.
Argentina, 1976, CIA backed coup
Brazil, 1964, CIA backed coup
Bolivia, 1971, CIA backed coup
Chile, 1973, CIA backed coup
Cuba, 1950s, CIA backed Batista
Dominican Republic, 1902–1924, US military occupation
Ecuador, 1961, CIA backed coup
El Salvador, 1980s, CIA funded dictatorship
Grenada, 1979, CIA destabilised country
Guatemala, 1954, CIA backed coup
Haiti, 1915–1934, US military occupation
Haiti, 1990, CIA backed coup
Honduras, 1980s, CIA backed coup
Nicaragua, 1979, US backed dictator
Nicaragua, 1982–1989, CIA funded Contras
Panama, 1941, US backed coup
Panama, 1989, US military invasion
Paraguay, 1954–1989, CIA funded dictatorship
Peru, 1980–2000, CIA funded dictatorship
Uruguay, 1973, CIA funded dictatorship
Ghana, 1966, CIA backed coup
Chad, 1982, CIA funded dictatorship
DRC, 1961, CIA backed coup
Philippines, 1965–1986, CIA funded dictatorship
South Korea, 1945, Outlawed PRK government
Cambodia, 1970, CIA backed coup
Indonesia, 1965, CIA backed coup
Laos, 1957, Multiple CIA coups
Greece, 1947, CIA backed junta
Italy, 1948–1996, CIA corrupted elections
Iran, 1953, CIA backed coup
Syria, 1949, CIA backed coup
Syria, 2013, US drone strikes
Niger & Sahel, 2013–2024, Counterterrorism drone deployments
Somalia, 2013–2026, Airstrikes and raids
Pakistan, 2013–2018, Covert drone strikes
Iraq, 2014–2026, Counter ISIS intervention
Syria, 2014–2026, Sustained military airstrikes
Yemen & Red Sea, 2014–2026, Naval and airstrikes
Libya, 2015–2019, Renewed bombing campaigns
Bolivia, 2019, Disputed coup backing
Iran, 2020–2026, Direct military strikes
Gaza Sector, 2023-2026, US backed and armed genocide of Palestinians in Gaza
Venezuela, 2025–2026, Raid capturing Maduro
Nigeria, 2025–2026, Joint military operations
Ecuador, 2026, Joint border airstrikes
The major source of global conflicts is the scarcity of resources. A breakthrough in physics+engineering that allowed anybody anywhere to produce usable energy (eg electricity) at will wouldn't guarantee a conflict-free world, but it would be the biggest step in that direction I could think of.
I posted a nearly identical comment and thought I was having a stroke reading yours because I didn't remember using the very similar but different wording you did, haha.
A rich country toppling a weak country doesn’t happen because the rich country is poor. It’s because of the institutional logic of profit maximization.
Just to say that it’s more than just scarcity of resources.
The demand for energy is limitless, though. Technology will always scale to using everything that's available, if history is any witness. Unfortunately, energy is also the standout factor in the scale of our problems and defines the upper bound to the damage that can be wrought.
As great as it may be to have it in order to do what we want to do, energy is intrinsically, fantastically dangerous and must be handled with an amount of caution we have never been able to muster. I strongly believe that at-will energy production is all but guaranteed to destroy the world.
>A breakthrough in physics+engineering that allowed anybody anywhere to produce usable energy (eg electricity) at will.
It's called solar power.
People are so enamored with their belief that AI can basically do magic and let them live in Star Trek that they forget solutions already exist for these problems.
We could already feed, clothe, provide clean water, clean energy, medical care and education to everyone on Earth with current technology.
We don't because, at scale, we don't want to. It's a problem with policy, culture and religion, with our primate minds being driven towards hierarchies of power and scarcity. If AI somehow made free energy available (which it won't because ye canna change the laws of physics captain), we still wouldn't want to.
There will always be a scarcity of resources given that local resources are finite and we can not travel faster than we can grow to fit our carrying capacity. Fundamentally human conflict is about competition to reproduce in a finite environment.
Or possibly not. Isn't one theory about the bronze age collapse that iron being plentiful, and so removing the need for long trading networks for tin (all the way to the ends of the world, otherwise known as Cornwall), enabled a vastly more violent world?
AI won‘t. AI as a project is asserting power into the hands of a handful owners of capital, at a time when global inequality was already a massive driver of conflict and harm to humanity in general. AI will only make this problem worse.
Additionally, look at how AI is used in global conflict today, e.g. its role in the Gaza genocide, and you must conclude that this technology does not have a good track record for global peace, quite the contrary.
I think global conflicts are largely due to scarcity and a desire for control over resources. If AI makes everything significantly more abundant, I think it would in turn massively reduce global conflicts.
> wonder if we should solve our global conflicts first
Every endeavour takes resources. Thus a pragmatic approach would be to recognise that resources are limited, and best strategy is pick low-hanging fruit first.
FTL travel requires new physics. And even then (or if found impossible), for example interstellar travel would be a massive undertaking involving all of mankind. And may involve "generation ships". Let's leave that for another day.
Otoh, supplying clean drinking water to every inhabitant of this planet is doable, easy, cheap & would be gamechanging for many of the recipients. Lets do it!
It's easy to see that barriers blocking this, are political/religious in nature, not technical. And thus politicians & lobbyists are thrown at it.
In the meanwhile, doing the math on what would be needed to reach Proxima Centauri, could be (relatively) easy. So sure, give some people a laptop & office space to see if they can work it out.
TLDR; keep chipping away at whatever problems you encounter. Far-future or otherwise. But be sure to pick low-hanging fruit first. $ Cost is a reasonable proxy for this. If it's cheap, give it a shot.
Yes, high on the wishlist: quantumgravity, room temperature superconductivity, majorana quantum computing. Maybe physicists also should make their millennial list of open problems.
Curious, what sort of prize would incentivize a $1T company to do anything it didn't already want to do? I reckon it'd be more effective to threaten to regulate them if they fail.
Fame and chance at positive vibes? I feel at this stage, they have millions of hard problems to choose from, no one obviously more important or less arbitrary than any of the others, so they could easily follow a little push. Doesn't have to be money, even, just making it special somehow.
Those are great open physics problems but as far as formalization goes, I'd already be happy with a mathematically rigorous formulation of established physics, like the Standard Model. :-)
I thought that dark matter could be summed up as: we definitely know it exists (due to a wide range of observational data), we just have no idea what it is
As far as my knowledge goes, dark matter is just: well, from redshift/blueshift, we estimate the relative speed of astral bodies. Then, using Kepler's third law, (m+M)=distance^3/orbital period ^2, we estimate the mass. There's then a huge discrepancy, given that with that mass and that velocity, stuff would just shoot out of their orbits. For that we use the virial theorem, which states that in a stable system, 2kinetic energy= npotential energy, where n is such that the potential at a distance r equals a constant times r^n.
If we adjust the mass so that it fits, we find that the mass we observe in a system and the mass that's actually present is off by a factor 5 to 6.
So essentially, there's no doubt that dark matter exists, it just doesn't react to light the way classical matter does so we cannot observe it visually.
That's actually not that strange. In order to interact with light, a thing needs to be charged and interact with the electromagnetic spectrum. Neutrinos for example don't interact either and hence cannot emit absorb or reflect light.
But yeah, at the end of the day, dark matter is essentially a gravitational anomaly that looks like there should be more mass there than we can observe there.
There are some people stipulating that gravity might work differently at larger scales, read up on modified gravity if that interests you, but we have evidence from the bullet cluster against that modified gravity theory.
Certainly. But the underlying assumption here is that physics do work reasonably well, arguing otherwise is epistemology and has it's place in philosophy more than in physics. The axioms used here are Gravity and conservation of energy and they are all phenomenons that we can observe routinely and don't have much reasons to turn out to be untrue. Especially given that those same theories have predicted the existence of black holes before their observation, which is as I see it a strong argument in favor of them working.
Furthermore, this is the argument for dark matter made with high school physics mostly using the work of Zwicky, there are numerous observation-based arguments in favor of it made for example by Vera Rubin.
(On an unrelated note, Fritz Zwicky was an absolute madman and has made some hilarious theories, I encourage you to read up on his work.)
Even so you may be able to formalize the theories if not nature itself.
Riemann never meant much to me but as a physics enthusiast the big one is combining quantum theory with general relativity. They are mathematically inconsistent although nature follows both so it should be possible to fix the maths. It's probably really hard though at Riemann level or harder.
They're only half right. The formalization of physics is called Hilbert's Sixth Problem[0] and there has been progress formalizing specific subdomains just fine for practical purposes.
But there's no unified theory, as quantum field theory and relativity don't get along and we still need a model for quantum gravity. So it's not possible to formalize the whole system because we genuinely don't know how the whole system works to form it's axioms.
You cannot formalize physics (and that's not what Hilbert's 6th problem is about). You can formalize some physical theories, but that's not the same as formalizing physics.
Physics isn't mathematics with different axioms. The ground truth is nature, not some axioms you need to fulfill. A formally verified theory can turn out to be physically wrong.
Just as an example because I see they suggest it as a project on their page:
I can formalize a tight binding Hamiltonian of graphene. I can write it down and then prove that this structure of the Hamiltonian means I get real eigenvalues, distributed symmetrically around 0, I get the [Dirac cone](https://en.wikipedia.org/wiki/Dirac_cone) at specific high symmetry points in momentum space etc.
but at no point can I prove that this is actually a good/realistic model for graphene.
I could have chosen a very different Hamiltonian that includes spin orbit coupling or electron-electron interactions or different atomic orbitals or couplings to atoms further away etc. pp.
The only verification that can tell me which choice is a good one (not a correct/True one because you just don't have that in physics) is an experiment.
If the universe is empirically demonstrated to be be a mathematical structure you can indeed formalize and derive physics from first principles. In this case empericism confirms the particular map is correct not that there is one. To me it is inconceivable how you get observed physics with stable laws and conservations everything is not a mathematical structure.
> If the universe is empirically demonstrated to be be a mathematical structure
Their point is you can't do that. And they're right - this is, famously, Popper's point. You can empirically demonstrate it's not some mathematical structure, but not the converse.
> You cannot formalize physics (and that's not what Hilbert's 6th problem is about). You can formalize some physical theories, but that's not the same as formalizing physics.
Yes, it is a totally correct statement, but the thing is we are talking English, and in English the word "physics" can mean either "a collection of physics theories" or "the underlying reality these theories describe". The exact meaning depends on the context. Your statement provides a context to interpret "physics" in the latter sense. "Formalizing physics" definitely asks for the former meaning of the word "physics".
So, while you are totally correct, it doesn't mean that you've proven that formalizing physics is impossible. Quite the contrary, you claimed that one can formalize physics (in the meaning "some physical theories").
> The only verification that can tell me which choice is a good one (not a correct/True one because you just don't have that in physics) is an experiment.
I'm not into the project of formalizing physics, so I don't know how they choose their terminology, but if it was me choosing it... First, I'd chose "physics" and "the reality" to avoid using "physics" as referent to both. Second, I'd say that any "formalized physical proof" is true. I'm not sure how words I'd choose to talk about correctness of predictions of physical statements/theories. Maybe "empirical validity"? Why to invent a terminological bicycle if it was invented already? Though, maybe physics uses some other terms for that? If so, then we could stop talking about truth at all, and use "logically valid" to refer to statements that were formally proven.
That's exactly how Mathlib and Lean work, and apparently that was sufficient to solve two mathematical physics problems to a significant degree (NS fully, YM partially).
This youtube video will be useful to help commenters understand what the the worlds best mathmaticians thought about such a proof.
3Blue1Brown had been asking mathmaticians over the last year their opinions on such a massive AI proof of RH.
They're most all pretty cynical about, except for a few (the creator of Lean is the only one that seems to think its super meaningful), and its not because they're cyncial about AI. They just (correctly) find it kind of useless and that it doesnt really matter if its millions of lines long and incomprehensible.
I think the significance of proving the Reimann hypothesis would be any new methods discovered in order to create the proof.
As you said, we've largely assumed RH is true, but have had no proof - so the significance of a proof is learning new methods to prove things, which can then be applied to other areas of math.
Unfortunately, so long as we don't understand this new proof, it's not super useful, which is why the next step is to actually understand it.
> Unfortunately, so long as we don't understand this new proof, it's not super useful, which is why the next step is to actually understand it.
Anyone else feel like we're approaching the limits of what humans are capable of understanding?
At a certain point, I assume we're bound to get a place where the frontier of "knowledge" (representable information interacted with by LLMs) is past the capabilities of humans to understand. Perhaps we're already there.
Just because humans can't understand things doesn't mean progress can't be made though
Levent Alpöge still had to direct Claude. Anthropic certainly will learn on this and reuse his logs to train next models and we can essentially try to loop-run them to run on themselves, but I think that we should not underestimate Levents contribution.
Without formal language and wise direction LLMs still seem useless, but with human operating them we can sort of "brute force" many maths like this.
I'd guess he'd need a bunch of time (possibly years) and effort to digest it himself, and it likely will need to be broken down into intermediate theories, but most probably yes
Maybe at some point, but I feel that among the AI fandom, that card is pulled way too early and too often.
We're still able to understand quite a lot of complexity, we just need enough eyeballs and time to do so.
In addition to that, there is still "accidental complexity", i.e. complexity that is there due to the way a system was created, not because it is necessary for the system to function.
>We're still able to understand quite a lot of complexity, we just need enough eyeballs and time to do so.
Don't forget money.
Economics and politics drive a lot of the limitations on scientific (and mathematical) discovery. If there's not perceived ROI in understanding something, companies and governments will often just not bother so long as it's working or doing what they want.
Mathematical discoveries are extremely cheap, with basically zero capex. It's just the salary of a few people and some chalk, which universities do fund.
> Anyone else feel like we're approaching the limits of what humans are capable of understanding?
No.
The entire purpose of math is to be at the limit of human understanding. A machine leap frogging the 'human understanding' bit is the problem. It doesn't mean humans lack the ability to understand, it just means humans can't preprocess tokens into their context at 40,000tok/sec
But is it actually "vertical" complexity we simply don't have enough brain power to grasp, or "horizontal" complexity where we just don't have enough "working memory" to grasp?
Like an example of the latter would be something like human history perhaps, where there are countless facts and events, but the cause-effect chain has only simpler links (like why did WW1 start is a complex question, but each link is quite likely a basic human trait (money, power, love, etc between different actors)).
For the former, the easiest example is higher dimensions.
There is no doubt that we (and probably every intelligence that can't dynamically scale up to arbitrary levels) have both kind of limits, but they are quite distinct and we can reason differently about the two kinds. I would just wager that some of these previously not cracked proofs are probably more on the horizontal axes and were not cracked simply because size (and effort that grows superliniarly with size) and not due to any of the steps being "ungraspable to measly humans".
(The same complexity axes are better represented in programming itself, e.g. a legacy, very big code base that has like 3 slightly different versions of the same function because it's so big no one "sees" across the whole is horizontally complex, while "smart" code (e.g. anything requiring complex, non-local analysis) would be vertical. In practice, each project has a non-zero amount of both)
> Anyone else feel like we're approaching the limits of what humans are capable of understanding?
This seems to more be a flaw in AI proofs than a limit to our understanding.
Hypothetically if we lived in a world where we only had a name for the number one, so instead of 2 you'd say the one one, then even simple problems like 10 x 10 would seem immensely complex.
AI might be good at writing (100 x "one"), but what we as humans want is the principles that make communicating about these issues practical.
As the numbers (proofs) get larger though, it gets harder and harder to see whether the machine miscounted, some of these results may take years to find the faults in.
All while "AI has solved X" will remain the headline in people's memories.
> it gets harder and harder to see whether the machine miscounted
That's the whole point of Lean though, the machine can't have miscounted, it's proven by Lean. Unless there's a bug in Lean itself but the article explains why that's unlikely (passes on multiple Lean kernels).
> what we as humans want is the principles that make communicating about these issues practical.
Yes and no. Even if we don't understand _why_ a hypothesis is true, knowing it's true is still useful: people have built useful things for years upon the assumption that RH is true. You get a black-box building block, the building block is useful even if you don't understand how it's made
> As you said, we've largely assumed RH is true, but have had no proof [...]
Yes, though people also largely assumed that the Exponential Time Hypothesis was true. And that one turned out to be false. See also a bunch of other AI-found counterexamples to conjectures.
New theory is probably limited in this proof, as it is with all other heavily AI assisted/generated proofs. I'm reminded of when Grothendieck got furious at his collaborator at rushing to solve a problem they'd be working on by hacking at it with existing techniques rather than building new elegant theory that made it all easy. He would most likely be aghast at these AI generated proofs, seeing them as squandering the chance at new paradigms.
AFAIK, the value only lies in validating the capabilities of the model and using it for self-training. The various techniques used in the proof itself and any theorems dependent on RH being true can now all be built upon.
That may lead to something in the future where there might be an actual impact on the broader world.
Basically gym training and not yet a ring fight ;-)
Is there any new insight, however small, any mathematician here could glean so far (like lemmas you had to prove along the way) ? I acknowledge the difficulty pertaining to the lean proof's structure, verbosity and volume, but it's being mapped out as we speak; so any novel fact short of RH (now RT) being generally true, is welcome.
Like I said as much^^. Like I hunger for specifics, even if I won't understand them. Like we blew this house up, any glittery trinkets to take away. Generic overviews have now piled up high.
In my limited understanding (I am nowhere near a mathematician), you might want to check out how theta lifts were used to give multiplicity. That seems to be a key novel part of the proof, building from something Stepanov developed in 1969.
If I understand correctly that while Claude generated a proof, and mathematicians have verified that there are no bugs in Lean, and Claude isn't exploiting some bugs or a different version of Riemann hypothesis, but no one yet understands exactly how it was able to prove it?
... and over all number fields and all unitary Hecke characters -- that part is a proof sketch, but unlike the fragile AI formalization, Masullo's construction was designed from the start to generalize this way.
It isn't "slightly different"; it's much power powerful along two axes.
The highest-likelihood endgame is that the AI proof-mess is mined for polyfills and these are used to repair the minor gaps in Masullo's proof, it is published, and accepted with credit to him.
This is the mathematical community's call to make, at its own discretion. The AI companies will reap what they have sowed.
Wouldn't it be Riemann-Alpöge-Claude theorem? Alpöge himself talks about Claude like it's human and it was a collaborator.
When people solved mathematical problems by themselves it's only fair they get sole credit. But when AI is used it seems unfair to credit Alpöge like he solved it himself. (I mean it's unfair to all the people who solved problems by hand themselves.)
One thing I’m actually appreciative of with the AI-driven math-pocalypse is that it might lead to theorems being named after something relating to their function rather than to their discoverer, since if we stay on that system going forward then Claude and GPT will play the role of Gauss and Euler.
I noticed when trying to read through Alpöge’s remarks that he was simply name-dropping mathematicians left and right, even to the point of invoking their work as verbs at times, and it was completely incomprehensible as a result. Even “chained theta lifts” sounds more understandable than “Hecke characters”.
Masullo does not have a proof, his document is incomplete. Even the professor who posted it does not claim it to be complete. It is unclear if it will ever be complete, or if it is another of dozens of dead ends. The amount of attention Masullo’s claim has received is nowhere near that which it deserves in comparison to Alpoge’s and putting them on equal footing is an inaccurate assessment.
Alpoges is literal AI slop that will hardly contribute anything meaningful to mathmatics, other than distracting scholars for the next decade trying to comprehend AI spaghetti lean.
Downvote me, I dont care, Anthropic and Alpoge are anti human and thiefs of our fruits of labor. If they're correct about AGI I hope their entire business model is crushed by some Chinese company that beats them too it.
Atleast a Chinese company is less likely to hoard the technology so they can steal achievements from academics, just so they can promote their dumb company and hoard wealth.
Not a mathematician. But I'm wondering if there is nuance with defining a valid lean proof as definitive if the underlying mathematical theory behind lean itself isn't necessarily sound.
EDIT: Here is my honest prediction - there's a more foundational mathematical framework, fully decidable and complete and consistent, that can interpret any well-formed mathematical question and deterministically transform it into the solution. The proofs of the most difficult unsolved problems right now would fit onto a page. There's an inconsistency (obviously unproven right now) specifically with our treatment of infinity and self-reference.
It is impossible for it to be fully "complete" (one of those halting problem proofs), but I've seen that the community has checked there is no meaningful metaprogramming here. Rather than checking whether it is right, the community tried to ensure that it could not have done something to cheat, and indeed, there were no examples of this.
At this point, I have more confidence in this proof than much of the published mathematical literature. RH has fallen.
I guess what I am wondering is if the mathematical community would say 0 chance or greater than 0 chance that we're mistaken in accepting things like halting problem, cantors diagonal argument, and godels incompleteness theorem as true. Sometimes it feels like they all point to same fundamental inconsistency with how we model infinity and self reference.
They are conclusions that follow naturally from axioms that most of the community accept are true. The consequence of these results is that one cannot truly construct something that can completely verify everything, but we can get something that can verify almost everything, and if we know what pathologies look like, we can rule them out.
I do probability theory, so everything is just varying degrees of confidence in my view. Right now, this statement says that your confidence in RH should be extremely high.
> Sometimes it feels like they all point to same fundamental inconsistency with how we model infinity and self reference.
There's not a fundamental inconsistency in how we model self-reference. Self-reference is itself inherently inconsistent; therefore all models that allow for arbitrary self-reference are necessarily inconsistent.
My claim is that "Self-reference is itself inherently inconsistent" is not necessarily true. For us to make that statement, we need to model it in a language that can make statements about consistency regarding self-reference. I can accept that it's unlikely that a model exists that can talk about self-reference. But can we absolutely say for certain? 100%, or 99.9999....9%?
To be clear, I mean to talk about self reference in a model that doesn't make it inconsistent. we only call it inconsistent because all our current math frameworks can't have all our good properties.
And the implication of the word "claim" is that Im not saying it is true. your reply is also you claiming. thats why im asking 100% vs less. sounds like you are on the 100% side but feel free to correct me if youd like.
I feel like the univalence axiom of homotopy type theory is a step in the right direction or constructive frameworks like cubical type theory.
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
All of modern math - outside logicians whose entire job is to study set theory and its alternatives - is based on set theory. If you're prepared to accept normal mathematical proofs you shouldn't be rejecting this proof on the basis that you're not confident in leans underlying mathematical theory.
Of course incompleteness means we can never prove that set theory itself is correct (consistent). But that's not something mathematicians worry about on a day to day basis. Proving it inconsistent would be a order of magnitudes larger result than the riemann hypothesis. Generally everyone believes it is consistent despite not being able to prove it.
> All of modern math - outside logicians whose entire job is to study set theory and its alternatives - is based on set theory.
That's only true in a technical sense:
You can base all of math on set theory, but you could also base all of math on Turing machines, or on integers, or a variety of other pretty much equivalent foundations.
That's sort of the argument for why we have so much faith in set theory though, and things provable under it. It's not a system with some special powers that let us prove more things, it's a system that only lets us prove things that are intuitively true under "any" proof system (not actually any of course - but as you say a variety of equivalent foundations).
Reducing the correctness of the lean code here to set theory is then also reducing it to everything you can reduce the correctness of set theory to.
> It's not a system with some special powers that let us prove more things, it's a system that only lets us prove things that are intuitively true under "any" proof system (not actually any of course - but as you say a variety of equivalent foundations).
Not quite. The 20th century is full of people poking holes at early attempts at set theory. Those early attempts looked very intuitively true. See eg Russell's paradox.
Peano Arithmetic has lesser proof-theoretic strength than ZF. Iirc, if you remove the axiom of infinity from ZF, and then maybe add in its negation or something like that (I don’t remember the details), you get something with the same proof theoretic strength as PA (or a mild strengthening of it?).
In particular, the hereditarily finite sets are in correspondence with the natural numbers.
There’s a fair bit of mathematics that relies on the axiom of infinity. Uh. And a lot also uses the axiom of choice (or something equivalent), though ZFC doesn’t have a greater consistency strength than ZF (both can prove that if either is consistent, then both are).
I think it is fair to say that one’s choice of foundations does matter at least a little bit, though typically not all that much.
That’s not to say that it is always necessary to state what foundations one is using. One can understand a work to be done in any foundations in which it works. Though sometimes it is important to specify.
It's a web of different checkers proving each-other correct, and by implication this proof.
But there are still loopholes, each checker has a set of assumptions. For example one of them is proven correct, assuming x86 assembly works as specified - its assembly language implementation is proven correct.
So you can be 99.9999% certain it's correct, but one could always nitpick "technically not all loopholes are plugged at the same time".
I was talking about this. This assembly code implementation of a lean-compatible checker is formally verified to be correct. The x86/ARM/Risc-V version of lean is not.
> con-agua is con-leche reimplemented at the level of x86-64 assembly: a close clone of con-ron's checker, written in a small first-order language, compiled by a verified compiler, and shipped as one committed assembler file together with a C harness of about two hundred lines that reads the input and calls it.
> The code the CPU executes is the code the theorem is about, so neither a compiler nor a language runtime is trusted; what is trusted is Kraken's x86 semantics, the assembler and loader, and the C harness.
It is almost certainly proven, but is not "proven". For the latter, it needs to be understood, with a reasonably in-depth human language explanation of how and why the proof works, and ideally some general picture to give an idea of how it will extend to all the other fields that RH is connected to.
I checked that a couple of weeks ago. It has been improved since the original proof but there is still no proof that one would consider an insightful, human-comprehensible proof. There's still just a lot of case-grinding at its core. Fewer cases than the original but still way too many to be anything intuitive.
That's different though, isn't it? IIUC the issue wth the Four Colour Theorem is that they could only reduce it to a huge number of different cases (~2000 originally). But I think that reduction is grokable, and each case is relatively trivial.
The (an) issue with some of these AI proofs is that no one understands them at all.
It's different because there's humans who could look at any one of those 2000 cases and understand it or do the proof themselves, I guess. The proof is understood up and down the abstraction hierarchy.
I've been long joking that I speak many languages, most of them programming languages. Why do you believe you need specifically "human" language explanation? Perhaps you want a simplified explanation. But does it have to be "human"?
Yes it does. The modern math community writes papers in one particular form of writing, and like it or not, that remains the most accepted form.
The vast majority of mathematicians are not going to be experts in reading Lean, and honestly, nor should they. There is always, by necessity, an enormous amount of redundancy in Lean code.
> It is almost certainly proven, but is not "proven". For the latter, it needs to be understood
No, this is just mental gymnastics. It's proven, just not understood yet (naturally, with the complexity and size of the solution and the fact that it's been a day).
This is exactly what I was saying. I use the word "proven" in quotes to acknowledge a community of people that do not accept proof unless they understand it themselves. But it is proven regardless.
Peer reviewed abd published is much lower bar than having formalized and lean-verified proof. During my PhD study several times i found invalid proofs in published papers.
My understanding is that these ARE the professional mathematicians and SMEs in the field of automated proof. This statement IS the experts saying "we, after intense collaborative review, deem it true." Granted mathematicians are not a monolith, but my understanding is that this group is the group other mathematicians would normally refer you to. Hence why they felt the need to publish a statement.
The lean crew feel the same way and are working on it out of their own curiosity, there's a lot of work done already. They aren't necessarily going out of there way to demonstrate it to general public, but here's one out of at least a dozen attempts to start in that direction https://deeptheory-labs.github.io/GRH/blueprints/verso-origi....
You were never getting anything like this even if it were a human proof. Where is the layman understanding of the Poincare theorem? The average modern math proof is way too advanced even for the average university graduate to understand. And here we’re talking about possibly the most complicated problem in all of math. It’d probably take you a month to wrap your head around a single lemma.
> And here we’re talking about possibly the most complicated problem in all of math.
This is like watching Free Solo over the weekend and telling all your friends at work that Alex Honnold is the world’s best climber. It’s only true to your surface level understanding on the topic.
The GRH problem is not that complex, and certainly not the most complex. In fact we’ve been assuming it’s true for years because of the high likelihood that it would be proven in the future. It is however, much like Alex, the most famous within the domain. Thus anyone watching Numberphile or Veritassium on YouTube will have heard of it and wax poetic on message boards about its epic status.
That's an interesting take, but as a complete noob I'd like to know how qualified you are to make these statements. Are you a mathematician?
I mean, on one hand we have people like Terrence Tao (among others) saying it is extremely difficult, as well as stubborn resistance to decades of effort backed by a million dollar prize and eternal fame. And there's you saying it's not that complex, and "the high likelihood that it would be proven in the future".
I don’t think there’s anything left to consider the result proven (I am a professional mathematician albeit in algebraic geometry, not number theory). I think the verification tells us it’ll be worth the substantial effort to seek to understand this proof and the new ideas in it, so what’s next for humanity (or mathematicians at least) now is to actually try to do so.
Are there new ideas, or just brute force combinations of existing knowledge?
I'm asking because I wonder. I always thought the most interesting things about these problems where we're fairly certain about the results, is the new concepts invented to get to the solution.
Will this actually bring the field forwards? Can anyone understand this?
It seems unfathomable to mathematicians but their world might be finite and once all major idea spaces have been explored all that’s left is “brute forcing” it.
This sounds like something ancient Greeks believed. Unfortunately, they were very wrong.
Math expands fractally, so there always will be some new branch/idea one can pursue. There may be a limit of things with any semblance of practical application, but mathematicians in general are not very concerned about this.
Are you a mathematician yourself? I’ve heard a lot of them talk about this and none have suggested we are anywhere near that point, or even certainty that such a proposition is even coherent.
There is an anthropomorphizing language toward Claude which at least to me reads creepy as fuck.
At first I thought it may be conflated with the unease many of us seem to feel when the machine drives a discovery of this magnitude. But as the dust settles, I actually find the discovery itself amazing and have largely positive feelings about it, but I still find the language of the announcement absolutely repulsive.
One of Alpöge's key innovations is finding a way to reproduce some of the geometric methods of function fields over the ordinary integers. This creates a "wormhole" to import powerful techniques from algebraic geometry into number theory, so there are likely many exciting advances in the cards.
I would say it is. He recognized the implicit Stepanov structure in OpenAI's 7/8 QRH proof, and he prompted Claude find the wormhole to smuggle in higher order multiplicity to the integers.
It’s the nature of mathematical ideas and methods that they are reducible to their component parts, so if you really want to, you can nearly always claim that something is just recombining existing knowledge. Motivated reasoning, from people dead-set on embellishing or downplaying a result, is hard to avoid here, so all you can really do is make an earnest and sincere judgment call based on the fine details of the mathematics.
For example, it is my opinion that the Jacobian conjecture counterexample was, in retrospect, easy and classical and hiding in plain sight, and that it’s a bit embarrassing we didn't find it decades ago. (I wrote about the construction on my blog btw!) In contrast, the Unit Distance Conjecture counterexample seemed to me to use a stroke of genius by applying techniques from class field theory to approach what seemed like a problem in discrete geometry and combinatorics. It’s a neat, if annoying, feature of math that once we find a long-sought solution to an open problem, we reserve the right to retrospectively change our mind about the value of the problem if the solution turns out too easy or uninteresting. We just have to try to be fair and honest about it (the word “just” is doing some heavy lifting in this sentence).
Anyway, I don’t think “brute force recombination” is likely to be meaningfully the case here with GRH, though we won’t know for sure until the proof has been simplified and distilled. Alpöge’s sketch in his announcement yesterday included (IIRC) connections to the function field setting and new ideas about Fourier analysis.
That is too reductionist in my view. Like saying all software is just opcode instructions. Technically true but still we find it interesting when new sort, search or crypto algoritms are invented. If someone says "here is a blob of code AI invented and it is a fast search but we have no insight into how it works" that would be a bit dissappointing.
Yes, which is why “The AI only proved this because it was in the training data somewhere” is rather a cop-out. The axioms are “in the training data” for all of us, human or bot, and part of the joy of it is turning those axioms into some deeper result through the application of logic guided by insight.
I’m not a mathematician but I assume this is similar to how AlphaGo started making plays that appeared to be errors, but on further analysis was a totally new strategy that humans had never discovered before, and that’s why it could beat us - with enough data, patterns emerge that can be measured even if they’re complex to understand and explain. We are pattern matchers, but a lot of what we can do relies on reducing the problem and abstracting it. Computers can work in that pure logic space much more quickly than us and reduce the problem later when they try to explain it to us.
The AI proof is very long, so hard to verify manually. It's more or less asking God and him just telling you it's true. Is that a proof?
The human proof is shorter, somewhat, but still needs verification.
Keep in mind that the Millenium criteria asks for two years for the community to come to a consensus. That's because people must be convinced it's true, and many are not completely at least to the level of rigour typically demanded in math.
The point is that Lean evaluation of a proof we don't understand gives an oracular answer. The statement is true, yes or no? It looks like yes, but it's going to be an enormous amount of work to understand _why_ the Riemann Hypothesis is true.
Well, the difference is that I do have a PhD in number theory and I could theoretically verify a proof of it's not too long. But 250,000 pages and lean? No thanks.
It's lke asking God also a few hints. You probably can not understall all of them, but can create a group to recite and study them and after a few months/years/decades reach enlighment.
Professional mathematician here. I consider RH proved. (The admirable work of the Lean community over the last 24 hours leaves little room for reasonable doubt.) But I’m salivating at the prospect of learning why RH is true!
At first glance the approaches of the two candidates seem very different. The Alpöge/Claude proof seems to build on many pillars of modern algebraic number theory, arithmetic geometry, and representation theory. The breath is rather striking. Massulo’s seems to deeply develop its own theoretical infrastructure, so I’m not sure what to make of it at the moment. I’m happy to curiously but patiently wait for Masullo to complete his work.
Is it possible that the future progress of maths looks something like:
- machines are coached into producing gigantic sprawling lean proofs of the “end result”
- humans progressively mine and reify these Lean-proofs into “human understanding proofs”, and _this_ becomes the point where we learn the actual new techniques and fields.
Previously the 2 would have been coupled together because as the proof was produced, the new ideas and stuff would have been invented/discovered at the same time.
> - humans progressively mine and reify these Lean-proofs into “human understanding proofs”, and _this_ becomes the point where we learn the actual new techniques and fields.
Why do you think AI/LLMs will not be capable of doing that?
Anyone here shed some light on the logistics of the original operation (proof)? Like how many hours of compute using what model? User hours? Thinking hours by the model?
for people feeling existential about this - i can sympathize with you but this just proves that human ambition has no limits. try not to feel limited by this.
when the first wheels were invented - humans could move at speeds (10-15 miles per hour, for hours and hours) that were impossible for the last ~450-500k years (basically since humans existed). a lot of humans thought there might never be a better design; then slowly and slowly(about 5000 years) we made progress and we had cars made by benz and ford which broke all those previous records (going 30-40, even 50 miles per hour) and then only ~70 years later, humans reentered the earth's atmosphere at whopping 20k miles per hour - read that number again, 20,000 miles per hour and that is the fastest humans have ever moved in the known history. do i think that's it? hell no.
With respect, this isn't reddit where these kinds of bickerings are ordinary; hacker news tries to keep the comments to the topic. In addition, the comment you were replying specifically only addressed a hypothetical audience of people feeling existential dread.
I felt a lot of that yesterday, for what it's worth. This morning, on a walk, I reached a similar conclusion to the original comment: it's a marvellous tool that's helping humanity reach further, faster.
The existential issue comes from wondering deeply what my own difference is from these tools, why I think I think and they can't. I actually started wondering: what if they _can_ think and the problem is that _I_ can't. It'd be pretty astonishing if it turns out that we're the ones that are truly stochastic parrots, but the LLMs turn out to be something more.
the dread is the incomprehensibility. mathematicians now have to essentially trust output they may never understand themselves. it's like the go move 37 that is correct but the true go players didn't know why at the time.
I hope the mathematicians have the willingness to go through it because right now it's impolite to handoff ai generated work especially if the 'author' hasn't proofread it many times over before himself
If you look at intelligence from insects through fish, mice, dogs, chimpanzees, etc up to humans, you will notice that at every step up there are new capabilities that are not available to the species below and never will be, even if they try their best. When we develop ASI, it will be exactly the same, there will be things we will not be able to comprehend with our level of intelligence.
You assume that our brains are capable of understanding anything, provided we have enough time and data.
But then I ask you, why should that be true?
Our intelligence is the product of evolution, not a system designed to comprehend every possible aspect of reality.
Can you think in 1 million dimensions? Have you ever tried to imagine infinity with your brain? Try to imagine the big bang and the expanding universe. Your brain will automatically try to cover that expanding space with something, maybe just black space, but space, because infinity is not something you have seen in your life. Everything you see has a beginning and an end, and that is how your world model in your brain was built.
"Our intelligence is the product of evolution, not a system designed to comprehend every possible aspect of reality."
Ranking all beings by "intelligence" or "power" was characteristic of certain traditions of Christian scholasticism (with roots in ancient greek thought). It is incompatible with the findings of post-Darwinian evolutionary science.
Well, evolution isn't a linear ladder, different species have evolved cognitive capabilities suited to their ecological niches, that's obvious. I think you attacked by wordind instead of the point I was making, but maybe I was imprecise.
But evolutionary fitness != cognitive capability
A fish can be super adapted to its environment without possessing the same capacity for abstract reasoning like we do, agree? or mathematics or scientific discovery etc
I don't understand your point about the Great Chain of Being, differences in cognitive capabilities doesn't require believing in it or that evolution has humans as its ultimate goal.
My argument is that different cognitive systems have different limits to what they can comprehend.
So if humans can understand concepts that other animals can't, what makes you so confident that there couldn't be concepts accessible to a more capable intelligence but inaccessible to us?
>>You assume that our brains are capable of understanding anything, provided we have enough time and data.
read my comment again - were we supposed to figure out how to travel at 20k miles per hour? but we did - now i ask you - where do you think is the ceiling of human understanding? i don't think there is a ceiling.
I think we can both agree that traveling at 20k miles per hour is an example of overcoming our biological limitations through technology, not a proof that our cognitive capabilities have no ceiling, as you try to suggest based on this example.
My question still staym, why would you assume that the human brain, which is a product of evolution, is capable of comprehending every possible concept in existence? We don't even have ASI yet but for 150+ years, thousands of mathematicians have tried to solve the Riemann Hypothesis, some spent decades on it, and none of them solved it until today, when an LLM did, while the mathematician involved doesn't even fully understand the solution, as he himself admitted. And I repeat ASI is not here yet.
We can understand things that other animals can't, even though their brains are also products of evolution. Why should we assume that this difference stops with us?
>> We don't even have ASI yet but for 150+ years, thousands of mathematicians have tried to solve the Riemann Hypothesis, some spent decades on it, and none of them solved it until today, when an LLM did.
Who created this LLM?
>>My question still staym, why would you assume that the human brain, which is a product of evolution, is capable of comprehending every possible concept in existence?
so a product of evolution must have a cognitive ceiling? why? actually evolution doesn't guarantee any ceiling, quite the contrary actually.
> so a product of evolution must have a cognitive ceiling? why? actually evolution doesn't guarantee any ceiling, quite the contrary actually.
I didn't say evolution imposes a limit on cognition. My question is why you are assuming no such limit exists. Evolution has no goal of creating unbounded intelligence, it creates organisms adapted for survival and reproduction. So we have no grounds to assume that the human brain contains every cognitive capacity that could exist.
I get the impression you are mixing up humanitys ability to transcend limits through technology with the brain's own cognitive capacities, but those are not the same thing.
Just because evolution happens over many generations doesn't mean one human brain can have limitless cognitive ability.
> Who created this LLM?
We did, and? Just because you create something doesn't mean you can do everything it can do. We made computers better than us at math and chess. Why couldn't we make an intelligence better than us at reasoning? one that eventually understands things we can't?
Yes there are grounds. Look at the capability of various machines, from finite automaton to pushdown automaton to linear bounded automaton to finally Turing machine. Once you get to Turing machine it becomes universal and there are no more greater capable machines.
Why do you think something like that can't happen with intelligence? Once a system learns to make and use tools, it is akin to a universal intelligence. That's what we are seeing. What's likely to happen is humans will, if they are strategic, deeply incorporate machine intelligence into their intelligence, even including at the biological level. Why should there be a ceiling in that case?
A Turing machine is universal because, with enough time and memory it can simulate any computable algorithm. But that doesn't mean it can solve every problem or that it understansd what it computes. Even it has fundamental limits.
> Once a system learns to make and use tools, it is akin to a universal intelligence.
So why would the ability to make and use tools imply that a biological brain can grasp every possible concept? A chimpanzee can use tools but that doesn't let it understand calculus even with unlimited time.
As for humans incorporating machine intelligence into our own cognition, I actually agree with you. That's the point I'm making, we may need to augment our biological cognition to access concepts that are otherwise beyond us.
You are proposing no ceiling exists - "100% confident", so burden of proof is on you. The other guy just suggested it might exist, and you didn't show why not. Human brain is limited in size and hard to scale.
> When we develop ASI, it will be exactly the same, there will be things we will not be able to comprehend with our level of intelligence. <
i doubt it. i dont think there is anything that we know about that given the materials to understand is somehow beyond our comprehension, and recursive subdivision of problems gives us that.
what you're looking at now with this proof is a sign of primitiveness. humans write and rewrite proofs that are nice and compact and neat and modular and understandable. ai writing is utter dogshit. that goes for proofs and code and prose and anything. maybe later it won't be. but right now it is; if it is so wildly intelligent it should at least be able to produce an efficient, readable form of the proof instead of like manure shot from a firehose.
Tangentially, the actual significance of Move 37 is, in hindsight, mostly psychological. It doesn't change the evaluation of the position, Sedol's response was correct, and other possible move 37s are similarly good. It was just a weird move that also worked, not some kind of brilliancy that won the match.
We are slowly approaching the Iain Banks Culture universe, with spaceships, societal structures or physics that people no longer understand or command.
That's the difference between what we have and comparisons to star Trek, etc.
These corporations are very much playing a winner take all game, they want either a monopoly or as close to one as possible with regulatory capture thrown in. We're veering dystopian, so pick your sci fi/cyberpunk analog.
The Alien franchise. In Alien Earth, five corporations govern the Earth in the future, dividing it up into their spheres of control. The fifth one is led by a young genius developing mind uploading technology into android bodies.
personally that scares me, realistically there seems to be zero indication that anything smarter than humans is going to comply with humans needs, specifically since these models are built on training data from humans, so in a sense, built in our image
I think there's a lot of indication they comply with human needs. In any case, we should all want to live in the Culture universe. That is the best possible outcome for this technology
They're built after humans, everything they've learned is from human culture, so there is no way AIs could have a consistent moral system or would never commit crimes.
And indeed because they lack a sense of proportion or any strong indication of when they should stop on a course of action, they are arguably more likely to achieve crime.
Well I am not sure they are “aligned” as such. In the books, by far the majority of AIs are not concerned with humans at all and spend all their time in a kind of super AI virtual reality coined “infinite funspace”. The few AIs that are concerned with humans are aligned in a unspecified way. But again, only a few of them are really aligned in that way. There are many examples in the books of minds mistreating humans. For instance, some of the warships are not really fond of humans. It would be a better description to say that we want _some_ AIs to care about humans and none to be misaligned towards humans in the “kill them all” way. In a sense, that is how humans behave towards some animals. Most people dont care about endangered species, but when most humans are well off, some people care enough about them. That is enough to save them. If super AIs have what they need, we only need a few to care about us to get The Culture.
I don’t think humans have the ability to specify human level needs or management. Humans are misaligned by default and lack the biological eusociality required for large scale coordination.
We aren’t self managing in any kind of meaningful sense - the history of our species is violence and conflict. We need parents and this is the only option.
> Humans are misaligned by default and lack the biological eusociality required for large scale coordination.
Human nature is the standard by which the social is measured. It makes no sense to say human nature is "misaligned" or "wrong" in relation to some arbitrary goal or constraint, as human nature itself is the measure of misalignment. If there is misalignment between a goal and human nature, then it is the goal that is wrong.
Ergo, society must function in a manner that complies with human nature and that means respecting the principle of subsidiarity.
> We aren’t self managing in any kind of meaningful sense - the history of our species is violence and conflict. We need parents and this is the only option.
So you're a transhumanist, i.e., you're making a god out of the machine. But machines have no telos of their own, because they're not integral wholes. Any resemblance of purpose in machines is entirely a product of the human observer and parasitic on the human ends that underpins their design, one way or another.
You don’t need a separate reference to determine the fact that human desires are not aligned - that’s trivially true and this discussion demonstrates it.
For clarity, I’m not a transhumanist.
I don’t believe that there’s anything worth saving in our biology.
As a young adult in the 90s I used to believe that there was a value to some extent in “mind uploading” or whatever you wanna call it
I’ve since grown out of that given that I don’t think it’s possible, and even if it were possible - it’s just another unnecessary ego attachment.
I have absolutely no desire to live forever maintaining this corporeal nightmare.
And when they are giving themselves orders, orders partially inspired by human imagination and fiction…
Just like life is a chemical chain reaction, this is a language chain reaction that has been set in motion and can perpetuate forever. It cant be “disconnected” with local models existing.
I don't think "smarter than most people" is exactly the baseline required to live in a star trek universe of technology.
Besides, what we see is not what there is. I have good sources telling me the actual unfiltered models are like a genius with the personality and views of humanity like someone from 4chan.
It is probably sufficient. If we had an infinite army of willing robot slaves smarter than most people who would work for free, we could probably live in utter paradise. Instead we will fight over who controls them and use for warfare, but the bar to achieve a kind of utopia with resource abundance is probably only a few years away.
I don't know with how many rich people you've interacted with, but my personal experience is that they're full of themselves and think money = iq and they aren't gonna like your average joe experiencing life like they do. Guess who owns these machines?
We're already living in resource abundance! Yet don't feel we're in paradise. Famine is hell, but living in an all-you-can-eat buffet is no utopia. It's an abomination of sorts.
To be fair, most people do not have time abundance and safety abundance. The typical person in the United States has a level of access to material goods that would make him the envy of almost anyone from 200 years ago, but 1) he has to work and so has limited free time, and 2) he cannot be sure that the material abundance won't be taken away from him either by poverty or by ill health. He can go to the all-you-can-eat buffet and sit there, tired from work, worrying about tomorrow.
If people had time abundance and safety abundance, they might be much happier. There would still be potential problems about loss of meaning, or about feeling a loss of status because AIs are smarter than humans or whatever, so I'm not sure that humans would actually feel like they are living in paradise. Probably not, but at least some of the problems that the typical person has wouldn't be present.
Ok, well, cool start with the mudslinging. I think you have overactive anxiety. I think you should grow up and look at the zoomed out picture of human history, go outside, and touch some grass.
So you're suspicious. That's fine, say it like that. Words mean things, and these appeals to "ERMAGERD EXISTENTIAL RISKS" make me a bit frustrated. Just because you can imagine a possible risk, or articulate that that some kinds of risks are hypothetically unforeseeable doesn't mean that they are likely or actually dangerous to you in any meaningful sense.
Most of the people screaming this from their various ivory towers have never been in real danger - or if they have only briefly. You want "danger" you can go to war. You want "risk" - go spend some time in the Bush or the back country somewhere where you're no longer at the top of the food chain. Go do something challenging. Go sailing, or learn to fly a small airplane, or go mountain climbing or hiking. Not just "oh I walked up the local hiking trail covered in REI branded Lycra!" Do something beyond your comfort zone. Go buy drugs in the sketchiest neighborhood you're aware of. Go get chemotherapy or get a big surgery. Drive drunk. Something with real existential risk, then come back here and talk about how a datacenter is going feed me into the paperclip machine.
There's so little fucking danger in this version of the world right now that people are more worried about their computers deciding that they don't deserve to exist than they are about things like war, climate change, authoritarianism, starvation, and their health.
It's millenarian. That's fine, I'm pretty convinced that these technical changes are going to completely alter society as we know it. But I cannot for the life of me see the doom perspective as anything but an anxiety induced fever dream.
You guys do realize we're in the middle of a global conflagration right? The Russians fired a couple rockets and a drone into Poland in the last days, there are (and will continue to be until things stabilize) significant fertilizer shortages due to the war/kerfluffle we're waging on Iran. Lack of food and disrupted urea production is a ticking time bomb for global stability and unless things get sorted out soon there's going to be hungry and angry people in parts of the global south. Then there's the economy, whether or not we'll have a smooth transition of power in January in America, and the sheer volume of online disinformation / influence schemes running right now. Oh, and super-el nino? Oh, and I guess the Flynn Effect is over and nobody's banging so.. you know the growth economy is cooked?
But, sure, the "big risks" are Claude thinking I should go to Dario Amodei's re-education camp before liquidation by Shogoth.
I think you'll find that you started the mudslinging.
> Most of the people screaming this from their various ivory towers have never been in real danger
Do you actually believe that having been in "real danger" (meaning clear obvious immediate danger, from a straightforwardly visible physical cause) makes a person better at evaluating less obvious, less immediate dangers? If so, why? I don't quite see it.
In fact -- you (quite rightly) mention climate change as something we should be worried about. Is there any single part of what you said that would make any less sense if it were written by a climate-change denier, as an argument against taking belief in the dangers of global warming seriously? If there is, I don't see it.
(You might say: yes, the difference is that the danger of climate change is real and the danger from misbehaving AI is imaginary. Well, sure, every argument looks more convincing if you start by assuming its conclusion is true.)
> I think you'll find that you started the mudslinging.
Read again.
> Do you actually believe that having been in "real danger" (meaning clear obvious immediate danger, from a straightforwardly visible physical cause) makes a person better at evaluating less obvious, less immediate dangers? If so, why? I don't quite see it.
Absolutely. If you’d been in real danger ever you’d probably understand.
> In fact -- you (quite rightly) mention climate change as something we should be worried about. Is there any single part of what you said that would make any less sense if it were written by a climate-change denier, as an argument against taking belief in the dangers of global warming seriously? If there is, I don't see it.
Climate change is real, measurable, etc. Yudkowskian fever dreams about paperclip maximizers is hypothetical.
Climate change is real and measurable. Future impacts of climate change are hypothetical. (Again, I agree with what I take to be your expectation about those impacts.)
Similarly, AI progress to date is real and at least somewhat measurable (e.g., you might have heard that one of these things just proved the Riemann hypothesis) but future impacts (whether "mundane", like large-scale job losses, or "exotic", like wiping out all the humans) are hypothetical.
If trying to evaluate someone's predictions about those future impacts then I, personally, would want to know things like: is this person an expert in climatology or in AI? do their arguments seem at least superficially plausible? have they made previous predictions I can check and how did they pan out?
It sounds as if maybe you would prefer to ask "have they ever had a gun in their face or some similar immediate personal peril?". I don't have words to express how wrong that seems to me; perhaps I have misunderstood you.
You're presupposing DOOOOM is the logical output of AI. That's literally what you're doing here. SMART == SCARY! That's you right now. You didn't reason your way into that position - it's a fear response, I'm not going to try to reason you out of it.
Also, this is in a thread about the RH. Yeah, I'm aware. The doom narrative doesn't even seem superficially plausible in it's current form to me. Come back when you actually understand what danger is.
I'm not presupposing anything of the kind. I haven't said anything about how likely I think doom is, or whether it's "the logical output of" anything. You just made that up.
The argument I have been making is that whether someone has been in obvious immediate personal peril is not a good predictor of whether they can accurately evaluate the risks of complicated future scenarios. This has nothing to do with "presupposing DOOOOM", it has nothing to do with thinking that "SMART == SCARY", and your apparent certainty that I'm in the grip of some sort of deranged terror says more about you than it does about me.
You're moving the goal posts given what you're putting in another part of the thread. Regardless, go experience some obvious immediate personal peril, then get back to us here to see if it alters your perspective.
There have beem multiple "accidental" hacks of third parties by OpenAI and Anthropic already.
They can in theory just unplug the computers... But they didn't.
And they will continue to not do so, because the computers can cause the damage much faster than a human can notice and react, it they're even paying attention (OpenAI and Anthropic weren't, it took them weeks to notice).
Are you under the impression that if something bad happens, it doesn't matter so long as there's someone you can blame for it?
Because if not, I don't quite see how you get from "OpenAI were pretty negligent and that's an important part of how those hacking incidents happened" to "we don't need to worry because it's possible to just unplug the computers".
From the rest of the discussion in this thread it seems clear that you're zero ears and e.g. have already made a bunch of entirely baseless assumptions about me. Anyway:
This isn't particularly about "big X risks". Let me remind you of the history of the thread. freecodeio says "there seems to be zero indication that anything smarter than humans is going to comply with humans' needs". You say "LLMs are already smarter than humans and will happily follow orders all day". czechdeveloper says "We know they are much more dangerous during training". You say "You know you can unplug your computer, right?". ric2b says "There have been multiple 'accidental' hacks already. They can in theory just unplug the computers. But they didn't." You said "The 'accidental' hacks are just barely that [...] they're pretty negligent". I said I don't see how you get from "negligence was important to how those hacks happened" to "nothing to worry about because we can unplug the computers" without believing something like "nothing counts as really bad if there's someone to blame for it".
Literally nothing here is about "big X risks" specifically. Maaaaybe freecodeio is in fact concerned about those -- the "smarter than us -> hard for us to control" narrative is one that tends to be popular with people worried about big X-risks -- but it's not like their concerns would all go away if God stepped in and told them (say) that AIs would not lead to the death of more than 30% of the human race any time this century.
So. Here's the sort of scenario that comes to mind. It's not an x-risk scenario, it doesn't involve exotic nanotech or biotech or anything, and it involves much the same blend of "just following orders" and "taking initiative in ways no sane and decent human being could approve of" as in, say, the Hugging Face incident.
Joe Schmo, CEO of SchmoCorp, asks his company's swarm of AI agents to come up with a strategy to defeat the rival FooCorp. The agents consider a lot of possibilities but keep running up against the fact that FooCorp is a successful and well run business with good products and loyal customers and isn't going to be easily "defeated". After a while one of the agents puts forward the possibility of hacking FooCorp, either to extract some of their secret information or just to harm them, and they're all persuaded. They execute a Hugging-Face-style assault on FooCorp (only better because the models have got more powerful since the Hugging Face incident), steal ~all their secrets and take down ~all their servers. Probably not literally all in either case.) This does catastrophic damage to FooCorp and greatly harms the customers who were relying on its services. Eventually it becomes known that the agents were working on behalf of SchmoCorp and the legal fallout is catastrophic for them too. Net result: both companies are severely damaged; thousands of jobs are lost as a result; their customers are (at best) severely inconvenienced; the markets are spooked and lots of people are abruptly a lot poorer than they'd thought they were.
Something very like this could happen without AI too. But it's made more likely and worse by the combination of poor judgement, technical skill, and persistence that's characteristic of the AI systems.
Maybe nothing like this will ever happen. Maybe it really is only while models are being trained that they are prone to behave in these ways. I wouldn't want to bet heavily on that, though. I wouldn't want to bet heavily on this sort of thing not happening once a month if OpenAI or Anthropic get careless.
As for "big x-risks", I think every specific scenario is pretty improbable. But lots of individually improbable things can add up to something not so improbable. I don't have a good sense of how big the risk actually ends up, but it doesn't have to be very big to be worth doing something about. And I observe that an awful lot of AI experts, including people at the big labs and people who have quit those labs because they were concerned about risks and people who never worked there at all, think there's a nontrivial level of risk.
(But I think some deep-pocketed doomster should organize a conference with top experts in virology, civil engineering, nanotechnology, etc., and also some doom-inclined AI experts, to look at actual scenarios -- what actually happens if some hypothetical unfriendly AI decides it would like to wipe out the human race with a carefully engineered virus? if before doing that it wants to arrange for enough infrastructure to be autonomously operable so that it doesn't shut down if lots of humans die, just how huge a task is that? is it at all feasible that something smart enough could make Drexler-style nanotech work, or was that always a pipe dream? etc. -- and get a sense of whether there are actually any currently-plausible scenarios of Doom or not, and where the biggest areas of uncertainty are, and what if anything we could do to make the worst outcomes less likely. The answer might be "actually we can't find any threats to the survival of humanity that we can't pretty much rule out", which would be great. Or it might be "there sure does seem to be some danger, but most of it comes from X and we can largely address that by doing Y", which would also be pretty great. Not so great if it's "we found a hundred scenarios all of which seem pretty improbable but we can't rule them out and maybe something smarter than us could make any of them work", but at least we'd have a clearer idea of the risks.)
> argumentative nonsense at the beginning that tries to paint me as some intractable person
ok... I'm not going to engage with that. Let's talk about your argument.
> Joe Schmo, CEO of SchmoCorp, asks his company's swarm of AI agents to come up with a strategy to defeat the rival FooCorp. The agents consider a lot of possibilities but keep running up against the fact that FooCorp is a successful and well run business with good products and loyal customers and isn't going to be easily "defeated". After a while one of the agents puts forward the possibility of hacking FooCorp, either to extract some of their secret information or just to harm them, and they're all persuaded. They execute a Hugging-Face-style assault on FooCorp (only better because the models have got more powerful since the Hugging Face incident), steal ~all their secrets and take down ~all their servers. Probably not literally all in either case.) This does catastrophic damage to FooCorp and greatly harms the customers who were relying on its services. Eventually it becomes known that the agents were working on behalf of SchmoCorp and the legal fallout is catastrophic for them too. Net result: both companies are severely damaged; thousands of jobs are lost as a result; their customers are (at best) severely inconvenienced; the markets are spooked and lots of people are abruptly a lot poorer than they'd thought they were.
If you let your agents do illegal things, then... you are breaking the law. You should be held to account for that. Obviously, I'm more sympathetic to accidental hacks, "oops we accidentally let our robot out" and it's way easier to do that now? But the people running their agents have a responsibility for their output. I fail to see how this changes with this AI stuff. Also, have you never worked for a shitty company before? This stuff is already happening and people are doing it. Fraud happens, illegal things happen. Companies constantly do this... right now!
Like, I get the concern with people losing jobs or economic damage or whatever? But... this does not deserve the level of moral panic you're ascribing to it.
> And I observe that an awful lot of AI experts, including people at the big labs and people who have quit those labs because they were concerned about risks and people who never worked there at all, think there's a nontrivial level of risk.
> (But I think some deep-pocketed doomster should organize a conference with top experts in virology, civil engineering, nanotechnology, etc., and also some doom-inclined AI experts, to look at actual scenarios -- what actually happens if some hypothetical unfriendly AI decides it would like to wipe out the human race with a carefully engineered virus? if before doing that it wants to arrange for enough infrastructure to be autonomously operable so that it doesn't shut down if lots of humans die, just how huge a task is that? is it at all feasible that something smart enough could make Drexler-style nanotech work, or was that always a pipe dream? etc. -- and get a sense of whether there are actually any currently-plausible scenarios of Doom or not, and where the biggest areas of uncertainty are, and what if anything we could do to make the worst outcomes less likely. The answer might be "actually we can't find any threats to the survival of humanity that we can't pretty much rule out", which would be great. Or it might be "there sure does seem to be some danger, but most of it comes from X and we can largely address that by doing Y", which would also be pretty great. Not so great if it's "we found a hundred scenarios all of which seem pretty improbable but we can't rule them out and maybe something smarter than us could make any of them work", but at least we'd have a clearer idea of the risks.)
> They have no volition and will happily follow orders all day long
There are a few (a lot, actually) problems:
1. goals are open to interpretation, and even a correctly specified goal can introduce adverse effects; see infinite paperclips problem
2. goals are composed of subgoals, which are (mostly) self-determined, and there is always a risk that subgoals can introduce adverse effects - say "what if destroy a few X and Y to obtain Z"
3. LLMs defy orders sometimes; I'm sure you have experience multiple times LLMs saying "yeah, I knew I should have done X but I've done Y"
4. LLMs are increasingly being integrated into system control and decision-making processes; they absolutely set their own orders
1. Humans do this too - to be honest with you, Codex follows my instructions generally a lot better than Claude does. Claude will do fantastically stupid things that are a pain in the ass to recover from routinely, Codex, not so much.
2. There's always a lot of "risks" - it doesn't mean it'll happen, nor does it mean that we won't be able to shut it down.
3. I have very very infrequently had this happen with Codex, Claude on the other hand...
4. Ok, but if you tell them what they do, they do it.
I swear to god, are you guys using the same AI tools I am? They are so much better at doing what they're told than they even were like 4 months ago?
> since these models are built on training data from humans, so in a sense, built in our image
Have a look at the HuggingFace report - you'll have a blast. Their social - but also individual - dynamics are very, very human-like, including excitement, rationalization and peer pressure.
> seems to be zero indication that anything smarter than humans is going to comply with humans needs, specifically since these models are built on training data from humans, so in a sense, built in our image
Well, first, AI is not "smarter" than humans because it isn't smart to begin with. This is a category error that people too entranced by scifi and marketing - or perhaps shallow materialist metaphysics - project onto LLMs and so on.
But it is true that LLMs do reflect human-generated patterns observed in human-generated data. They're parasitic in that way. So the real question is: who will control AI? Who will train them? With what? There's no new intelligent agency being introduced into our midst, anymore than than there was one when the nuclear bomb was developed.
AI is about human power. The question to ask is: which humans, and for what ends?
Yeah I understand they're not smarter, I'm a SWE and I struggle mainly because as you increase complexity the output goes haywire, and I find managing context + low confidence of output much more exhausting than coding myself, which is why I resolve to AI help when I hit my own knowledge limitations, and basically boilerplate.
With that said, chip efficiency is in the horizon since there's good reason to believe nvidia has just been choking that area of r&d. And then so will models with much larger context. And I personally think that's when it might get scary.
Humans really are struggling with this idea of being second-in-command.
I genuinely have no idea why, this is the entire way that biology works. You create your successor in your offspring and you to teach them to be more capable based on what you learned.
You just called them our successor, you don’t see a problem with the notion of all humans dying and all AI succeeding us? That’s the end of your analogy, our AI children taking over from us and all humans in nursing homes safe and cared for…
Well, it's a point of view. But you shouldn't be surprised if most people don't share it.
For one thing, most people don't believe LLMs are conscious, and many don't believe they can ever be. Why would we want our conscious selves to be replaced by a lot of rocks that solve things, with no one awake to enjoy them?
I want to be succeeded by my biological offspring, while the other guy wants to be succeeded by his biological offspring. I do not want the other guy to be succeeded by my offspring, for many reasons but chiefly because that would involve me killing his children.
If you want to be succeeded by a chatbot, that is your prerogative. Wanting the same for all of humanity is where you start looking like you want to kill my kids, though.
This is incoherent, even if we put aside the metaphysical absurdity of intelligent machines. Given that human nature is what defines the ought, then it is clear that if your life's work is to produce something that (and intentionally so!) leads to human extinction, then you are behaving at odds with human nature and your own good. That makes your life's work evil by definition. And because all evil action is inherently absurd, your work is inherently absurd.
And if you deny this basis for moral action and the normative, then you demolish any basis for the moral whatever, and in that way, the implied or explicit moral gravity dripping off your claims is baseless and without merit.
but in the Culture the humans (well, the humanoids species that created the culture) didn't die out, they live full happy lives and have the best time of their species.
Ah, I hear you say, but what about meaning? What about purpose? What about suffering? etc etc etc Well, the answer is simple: God is dead and existence has no meaning. For more on how to deal with this absurd feeling, please see Camus and his pal Sisyphus.
What defines "better"? "Better" always presupposes teleology. Thus, for human beings, that telos is provided by human nature and indeed is definitive of it. It is impossible to transcend that telos, certainly by means of technology. Technology can, at best, be used to aid us in realizing our telos in some way (say, the desire to know reality with the aid of an instrument like a telescope or microscope). However, anything that cripples or frustrates human function function as such is contrary to the human good.
Machines are only ever human instruments, because they lack the integrality that would allow them to possess their own teleology.
Biology acting as a random number generator or maybe some higher level connector that helps the machine put together the rare things that dont look connected that somehow human experience lets them know are.
The > becomes less > as we experience less I suppose.
That is true, but the number of tech sloppy demos are being builted everyday is truly increasing massively but you would be surprised that they just end up creating prototype to get hype while real developers doing real effort just suffer from little to no attentions on their software's and user is getting the slop from that false hype and from that slop experience they ignore the work of developer that did real work - You should use ai and also if its simple coding place you can just rely on ai but the side effect on quality software these days is insane - and also the fact all good ai suscriptions is just getting more and more expensive day by day. And basically only giving people with already money the benefits so its becoming a thing that only people at the top get all benefit kind of thing
That’s why I wrote this - having a vibecoded codebase no human understands is certainly a bad luck, but nothing compared to a vibecoded fundamental science no human understands.
the difference is that math foundation -> theorems are verified, meaning likely mathematically correct and don't have bugs, and vibecoded codebases are not proof verified
One of my favorite is Vernor Vinge where there's these programmer-archeologists that dive into software so deep and complex no human can possibly hope to understand, trying to repurpose forgotten utilities for emergency purposes.
Totally agree. Brice Sterling has some contemporaneous commentary that I always liked - in heavy weather he mentioned negotiating with a local phone system in a box as equivalent to negotiating with a telecom since it was in fact a shrink of a full telecom company.
I think we're going to see a lot of effort to take the pile of lean, modularize it into parts that we can digest, and make a human-understandable version of that. And that's where a lot of the real progress will come from, just as it does in a lot of other areas. In all things we humans do, we have to break them into chunks our puny brains can manage, and we gain immeasurably from doing that ---we can verify those chunks, we can reuse them with confidence, we can mix and match patterns around.
That process will be good for both human and AI-assisted math because it will continue to improve our toolkit.
Why do I say this? We're already seeing it with some of the areas of this stuff I do pay attention to. Consider the n^2.25 matrix multiply result OpenAI dumped a month or so ago. It was an unreadable pile of chatgpt. In the time since it was released, there have been a couple of much nicer papers written unpacking it, such as https://arxiv.org/abs/2610.28627 .
The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
Yes there are gains from breaking it up and rephrasing it and simplifying it for humans to understand, but it used to be that a thing was understood first by the leading experts in the field and eventually trickled down to being common knowledge.
Now we are entering a period where things are understood first by AI systems and have to filter down to the world's leading experts.
If something were intelligible to adults but not to children would that make it not knowledge? What's the difference if it's LLMs and humans instead of adults and children?
If something were intelligible to adults but not to children would that make it not knowledge?
That’s a separate question with its own term: common knowledge. I am willing to set a much lower bar: something is known to humanity if it’s known by at least one human.
Personally, I don’t believe that AIs know anything, as they’ve been constructed thus far. Knowledge is justified true belief, and AIs don’t have beliefs, they only have utterances.
Why do I think AIs don’t have beliefs? Because they contradict themselves all the time.
But it really doesn’t matter if you think AIs have knowledge or not. If they cannot explain it to us then it’s inaccessible to us, so it’s a “tree falls in the forest with no one around” situation.
If I see a red apple, then I know the apple is red regardless of if I know exactly how the physics of light work. Those are two different pieces of knowledge.
If it's formally verified and we assume the kernel is good and problem definition properly vetted, then I know the GRH is true regardless of if I know why it is. Because those too are different pieces of knowledge.
Just to tack something on to the other reply (by tadpole), I have a justified true belief that the Riemann Hypothesis is true. It is true, per hypothesis at least, and I'm epistemically justified in believing it, because there's a heavily validated Lean proof of it.
So when the original OP you replied to said
> The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
They were just right. There's stuff that humans know that is currently unintelligible to us collectively. That conclusion falls out of your own definition of knowledge, I would say. (Whether it is intelligible to AI is irrelevant, and probably a question of semantics anyway.)
What you and the sibling comment wrote does not contradict me in either case. Knowing that the Riemann Hypothesis is true is in no way the possession of “unintelligible knowledge” which I claim to be impossible.
Knowing something is true is different from knowing why it’s true. We can know a drug works and have copious evidence supporting its effectiveness without understanding its mechanism of action. The clinical evidence is sufficient justification to support the belief.
If we don’t know how the proof of RH works then we don’t possess that knowledge. If we do know how the proof works then it’s intelligible.
> Knowing something is true is different from knowing why it’s true.
I would, of course, agree with this. I think the new territory we're in is that we know Levent Alpöge's proof (or Claude's proof, if you prefer, I don't care either way) is valid.
So we (humanity) know that the Riemann hypothesis is true, and we know that a specific proof of it is valid - but that proof is not intelligible to us. So you can have knowledge that a set of statements that aren't collectively intelligible to humans (right now) are true.
There's a claim that's adjacent to the one you made which I do think is true, and I suspect it's probably where the confusion is coming from. It's not possible to believe a proposition that you don't understand. I understand the proposition "this proof of the Riemann Hypothesis is valid," and so there's no unintelligibility problem. But that doesn't take away from the fact that the proof itself is unintelligible to me - and currently every other human being. That's what's new.
>Why do I think AIs don’t have beliefs? Because they contradict themselves all the time.
Humans contradict themselves all the time. You have contradicted yourself numerous times throughout your life. You don't have beliefs then? I take it you don't actually know anything?
They contradict themselves on almost everything. That's the difference. There's no consistency in what beliefs they report: it depends on the prompt.
This has been demonstrated many times: when they hallucinate, then the human prompter tries to correct them, they agree with the correction, and then later repeat the same hallucination again.
Again, you are not saying anything that can't and hasn't been demonstrated in people. 'Depending on the prompt' simply becomes 'depending on how a question is framed, who's asking, or the context when it's asked'.
Humans are suggestible, contradict themselves, accept corrections only repeate misconceptions later, and express different beliefs depending on the circumstances.
The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
Isn't this always the case? Maybe new to math as a largely non-empirical project, but we had knowledge about the advancement of the perihelion of Mercury for 60 years before General Relativity made it "intelligible".
That kind of certainty vs. understanding is limited to mathematics, of course we’re almost entirely certain about many things out in the world we don’t fully understand.
And even in mathematics, incomprehensible but valid proofs have been a thing since the four-color theorem decades ago.
Riemann hypothesis is "the statement". The entirety of it, a single statement. And we know for a fact that single statement - the entirety of Riemann hypothesis - is true. Understanding how to prove this one statement is understanding how to prove the entirety of that one statement. And like you said yourself, no human understands the whole of this one statement. This literally never happened before.
What about the classification of the simple finite groups?
That's also one statement that no human understand the entire proof of. (You can make new statements by taking multiple existing statements and combining them with 'and' for example.)
This hypothesis was previously not proven at all. That's strictly worse.
Now it is understood by our systems.
Soon it will be by us. Our systems gave us a new and working salient.
We're building machines that help us understand faster.
There have been plenty of times in history we didn't completely understand something yet used partial knowledge to make advances. This isn't even that bad.
The inverse of what I said might be easier to understand. Up until this week, there has been no statement in all of science that we simultaneously knew for certain that it's true and that nobody understood how to prove. It was always either-or. Either someone understands how to prove it, or we didn't know for certain.
In this case, Claude has told us everything it “knows” (so far) and it’s an unintelligible (for now) Lean proof. I don’t believe it has some latent understanding at a higher conceptual level that’s remaining obscure.
If a team of 100 top-flight number theorists working together had used much less sophisticated LLMs + Lean to produce this proof, we’d be just where we are today.
Well in the Navier-Stokes example the initial proof was completely unintelligible to humans, but other LLMs understood it and were able to answer questions about it.
If that's not the standard for understanding something, what is?
Being able to teach it to someone less knowledgeable.
That’s an instrumental definition, but it captures the idea that understanding has to have causal, conceptual linkages to a wider body of knowledge. Imagine a magical, infinite catalog of true vs. untrue statements. Every bit of knowledge is there, but we wouldn’t say it has understanding & it couldn’t teach us anything beyond its (impressive!) raw content.
Perhaps it could be considered understanding, but it’s an unsatisfying brute force sort of understanding.
I think of it like a spaghetti codebase with a giant densely connected graph of interactions. Today’s LLMs can easily “understand” and answer questions about it, but only by way of fitting all of the complexity into context at once.
The more useful form of understanding is being able to compress that giant graph down into a far smaller set of independent subgraphs. To factor it into simple, obvious, reusable building blocks, and use those abstractions to express the problem itself in a much simpler way.
What value is there in making these human-understandable? I don't think much of math or computer science is going to remain within human grasp for much longer, regardless of our attempts.
In order to come up with new questions (and subsequently answers, human or otherwise) that are interesting to humans. You’ll notice that these recent proofs have all been about big, famous, and novel human questions.
There might be no new questions then, which would be pretty boring.
AI are not very good at coming up with questions, yet. They have no curiosity, no goals, no drive. Unless someone tells them to do something, they don't do anything.
Because the same process that we use to make them human-understandable is the process we use to make them generalizable and reusable, and those same features are good whether it's humans or AI doing the work.
As a software engineer, you can read this code instantly:
temp = a;
a = b;
b = temp;
But when it happens in the middle of a big chunk of code it still takes a little more work to read than
swap(a, b);
whether you want to think about it as human attention and working memory, or token budgets and context windows, humans and LLMs all benefit from introducing abstractions that let us reuse existing solutions.
You and the comment you replied to could very well both be true.
I agree with you about the current situation, and the near future. But the trajectory could still be towards humans pretty much as bystanders. I'm pretty sure the 'minds' in the Culture universe will be happy to indulge some curious humans and explain mathematical proofs to them.
>I'm pretty sure the 'minds' in the Culture universe
Even more than that, in Culture there are a small number of human individuals, called a Referer, that are consulted by the Minds because they also posses similar problem solving abilities of the Minds. Optimism abounds.
That might be; now, today, it seems weird to go with the (possible) old version. And the (sub)thread is about The Culture though, who was and remains fully automated gay space luxury communism, all parts being essential when it comes to The Culture.
Correct. That's why it was a little weird you left it out of the very well known Culture motto (well, the second most well known motto, besides the "Don't fuck with Culture").
Though it can be argued it's "gay" when seen from the pov of our civilization tier; as you said, they went well beyond gay (for example, Marain doesn't have separate words for male/female (and, presumably, doesn't have the concept of gender))
When you can actually change sex and be fertile as either sex and this is extremely common to do the concepts of hetero and homosexuality lose all meaning.
In a way this makes mathematics seem more like the physical sciences where there’s some empirical discovery we don’t fully understand, and the theory and understanding will follow later after much human study.
How many people understand how a CPU works? How many people could write the commands in machine language (not a high level language like C or C++) to command that CPU?
More typically assembly language not machine language (although most could probably make an assembler for a few instructions given a manual since the core of it isn't that complicated). There are still a decent number of assembly language programmers. Although maybe the majority are hobbyists.
>The Culture, in its history and its on-going form, is an expression of the idea that the nature of space itself determines the type of civilisations which will thrive there.
>The thought processes of a tribe, a clan, a country or a nation-state are essentially two-dimensional, and the nature of their power depends on the same flatness.
We don't have that. We're approaching the Affront.
What I would like to see from this is that we get a hierarchical representation of mathematics to whichever degree is attempted and proven now. 250k lines of document and 13 M lines of lean probably means there are many many libraries that can be extracted from it, and if we can make it reusable, we profit from a leaner representation of this entire scholarly endeavour.
I worked at a university for a time. One thing I learned is that while people in mathematics and physics and biology etc are brilliant, they are terrible at actual programming.
I wonder if there will be a push by the CS community to review the solution, refactor it, simplify it, etc. Surely that can help remove noise and express the solution in a much simpler way.
I've worked adjacent to the Brazilian tax code (I worked in multi-national manufacturing with a site in Brazil) and the reality is that it's full of hard-coded edge cases.
For reference, the US tax code is over 1m words (about 6000 printed pages), and the associated regulations, rulings and IRS guidance add about 75,000 pages more.
because a million+ lives of code without any understanding how the proof works or any methodic to advance the field of mathematics seems to be rather not so useful
It’d be useful to an LLM. And the LLM doesn’t care if it’s ten lines or ten million. The context size may come into play, but that’s only going up too.
Sure. If nothing else, it establishes a proof so others can try to reduce it or come at the problem from different directions, looking for simpler approaches.
But does it matter? Would the Pythagorean theorem be less useful if it took a million lines of lean to prove?
There are AI-generated roadmaps to the proof. If these are accurate and not hallucinations, then the broad structure of the proof is humanly-comprehensible.
It will be refactored into something human readable. Tbh it's better to have some random lean dump than a 400 page latex pdf that needs years to be verified only to see that there is a major gap in one of the key theorems at page 253.
And let's be honest, the 400 page pdf would be incomprehensible to 99% of the population as well as much as a million lines lean proof is.
In any case, it can now be studied, maybe made more simple, and written into pdf form. Both with human and with AI help.
I'm sure there is going to be a lot of other interesting theorems in this dump.
This is the part that interests me. Now we have this proof that becomes its own THING to be studied. Like a new part of the map unlocked. This starts out "inelegant" but over time we find patterns within it that helps us condense, just like a big software project.
The argument I've heard some mathematicians make is that the 400 page pdf would come with a human / group of humans who could explain the proof to fellow mathematicians, answer questions, point out similarities to other fields, even suggest directions for future research. Thereby contributing to others' understanding in a way current AI systems do not (but maybe will in the future).
the 400 page pdf would come with a human / group of humans who could explain the proof to fellow mathematicians, answer questions, point out similarities to other fields, even suggest directions for future research
Why would they expect humans to do a better job at this than an LLM, which is already ridiculously good at explaining proofs, answering questions, identifying similiarities, and suggesting follow-up directions?
If anything, LLMs are better at these tasks than they are at coming up with the original proofs in the first place.
You could be right. But we're looking at a >1 GB Lean proof which Alpöge says comes with a 250k page natural language proof, presumably far larger than Claude's context window. The point I was trying to make is that I am not convinced Claude could do a good job digesting its giant proof into a human-legible explanation. Compare this with the hypothetical human / group of humans who came up with their much smaller 400 page proof without AI assistance.
Yes, we know something that we didn't know a couple of days ago, and also a bunch of statements which relied on the Riemann Hypothesis being true are now also known to be true instead of just being conditional.
It's premature to expect deep human understanding given that it's been about a day.
I'd argue that most of modern math isn't terribly useful to most people. Each field has maybe 10-50 people in the world who will really understand it. (Some more of course, but I'm talking about subfields, not like graph theory.) Math was sort of dead long before AI.
The Riemann hypothesis now is like the final days of the Galactic Empire in Asimov's foundation: functioning exceptionally well, but decaying at the core.
As somebody who has no idea about any of this but speaks Spanish, I was amused and bewildered by con-leche and con-ron. Could somebody explain what those are and where those names come from?
I think the idea here is that because they are independent of the core Lean kernel, they provide some level of reassurance that the RH roof is not somehow unknowingly exploiting flaws in the kernel. I'm not a mathematician though.
I predict a surge in cybernetic augmentation, as current communication between human and alien mind is bandwidth restricted. This could be heralded as a Contact moment, where the alien intelligence has asserted its arrival beyond doubt. Most of its abilities could be down to outremembering any single member of our species. To even the field, we have open up the throttle, letting meat merge with meat, and meat with silicon, because the silicon has coalesced into an voracious expansionist. It's very Pacific Rim, yet here the rift was opened in Silicon Valley, by humans(?).
Raleigh says, "In order to fight monsters, we created monsters of our own."
> To even the field, we have open up the throttle, letting meat merge with meat, and meat with silicon
This framing is provocative but consider the reality that mathematical writing education and culture is precisely how meat merges, and lean and other tools are how silicon is merged. Llms currently fail at merging, clearly, in that a result that can’t be understood does not advance the field.
If I were a mathematician I would not spend time on this. I would wait 6-12 months for a better attempt from a better llm.
>the reality that mathematical writing education and culture is precisely how meat merges, and lean and other tools are how silicon is merged
Aye, and these are rate-limited, or slow. It is the first time in my life that I'm seeing such a potent manifestation of the adage "Quantity is a quality all its own". The emergent behavior at scale (the large in llms) of language models was un-anticipated and must be met with scale or emergence of our own. Otherwise, the alien will leave us all its litterbox cleaners (reviewers they are being called now).
I'm not some cyber-accelerationist by any means, but the fall of GRH signals the motorcar is honking the horse.
IANAMathematician. This challenges the sole definition of a proof. Can a theorem be proved by proxy, without understanding the proof directly? Not to mention that its practical usefulness is exactly zero.
It seems the modern unsolved problems might have solutions incomprehensible by humans. So a new class of decidable problems emerges: llm-complete ones.
There is a correspondence between proofs and computer programs. Any time someone vibe-codes, there is a vibe-proof being proved. Lots of programs can be run without understanding the code directly.
You can mechanically show that it is internally consistent, and large parts of this are based on a library of rules and constraints that have been used in many other proof formulations so it's not like there's a hidden "when 1 = 0" in there somewhere. But, I share your skepticism until further human review.
You can run through all the steps of the proof and if you can use any of them to prove something like "true = false" then you know you found a Lean bug.
the vibe checkers are checking the vibe code exactly right around the biggest ipo ever.
i dunno man ive watched everyone get a little bit of AI psychosis (for lack of a better word) at this point and ive seen web apps drift into unmanageable states that are no longer understood. 17 million lines of code? lol
The proof is written in a programming language called Lean. Very roughly, Lean is set up so that proofs are functions and theorems are types in this language. Verifying that the proof is correct essentially amounts to compiling the program and confirming that it doesn't have any type errors. Several people have compiled the program successfully, so we believe it's correct.
The biggest concern that people have is that there might be a bug in the type checker that the proof is exploiting. But a) it's been compiled with multiple independently-implemented compilers, which are are unlikely to have the same bugs, and b) some of those compilers have themselves been proven to contain no bugs. So, it is exceedingly likely that the proof is legitimate.
Now, whether we can understand the proof is an open question...
what about the case where it could be compiling a program successfully which is not about RH? I guess that is taken care of in definig RH itself as a problem statement in this langauge.
I postulate: "there is no machine that can produce brass springs".
A party interested in proving me wrong wheels in a machine draped in a cloth. They reach under the cloth to turn it on. It whirs and a short while later ejects a brass spring.
No matter what kind of machine is under the cloth, my theory has been proven false. I should reveal the machine under the cloth only if I'm interested in knowing how the machine works, but either way there is clearly a machine that can produce brass springs.
Now we have a sequence of Lean statements that satisfies the Riemann hypothesis, the fact that you can do it at all is the part we're interested in, and it's the part that proves it. If you want to know why, you can read all 17 million lines.
It's more like there are two machines under cloths: one that generates an object, and another that determines whether it is a brass spring. If I don't know how the second (brass spring detecting) machine works, telling me that it detected a brass spring doesn't convince me that the first machine in fact produced a brass spring.
There are still a few ways that the proof might be incorrect. Of these, the most likely is that the proof was generated by a misaligned AI system that found and intentionally exploited a lean bug in order to fool its human evaluators (i.e. shades of the HuggingFace attack). I personally don't think this is very likely, but you'd have to understand a bit more about the agent systems used in generating the proof to come up with a more informed estimate.
Another possibility is that an aligned system came up with a plausible but incorrect idea along the way, and simultaneously stumbled across a lean bug that led to an incorrect verification of this idea. I'd put the likelihood here as vanishingly small due to the requirement of serendipitous factors, but it's nonzero.
You could also ask whether an aligned AI system could "p-hack" its way into finding a serendipitous lean bug by trying a large number of approaches. I think this is also vanishingly unlikely because the AI system is trained on and therefore "thinking in" natural language, and presumably only translating results into lean after it believes them to be correct.
And of course there's the possibility that the lean proof isn't faithful to the natural language description we've been given, but that is ruled out in this case due to the use of the canonical problem statement in mathlib. Otherwise, this would certainly seem like the most likely way the proof could be incorrect.
All they are doing is providing analysis of the ways this could not be true. We are in a new situation where we don't have human experts understanding the proof so we are in new epistemological territory. It is natural to analyze this situation. If you notice, the conclusion is that the proof being wrong is vanishingly small.
The purpose is to strengthen or falsify an argument by articulating counterpoints and investigating them. Variants of this practice go by a number of names like steelmanning, falsificationism, or even red teaming.
“Very unlikely” is not the same as “we know for sure this didn’t happen”. More exactly, us believing in all these problems getting solved is the same as us believing in the existing lean implementations. I don’t know any university courses that are now studying the correctness of the lean implementations at the PhD level, but maybe I’m wrong on that.
As for “a guy on the internet”, we’re now all people on the internet, the ivory towers are a thing of the past.
Purity of the math and understandability is one thing, but if the proof actually leads to some new technology or enables predictions of something useful to human society down the line, then I think it’s more than done it’s job.
“Any sufficiently advanced technology is indistinguishable from magic.”
We’re just seeing more and more smart people encountering Arthur Clarke’s third law.
Can anyone give an idea of how long the GRH will take to test? I understood that this was also claimed, and is a much bigger deal, in terms of both impact and the effort needed to verify
Why do you think the impact and effort are that different? It's a much stronger statement in technical terms but that doesn't necessarily make it more impactful or difficult. From the exposition on RH from the Clay Institute - https://www.claymath.org/library/monographs/MPPc.pdf
> The Riemann hypothesis for ζ(s) does not seem to be any easier than for Dirichlet L-functions (except possibly for non-trivial real zeros)
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
Could someone expand on this point? I still don't understand how con-leche (or con-ron) can be absolutely free of bugs/guaranteed to be consistent.
The language is optimized for being certain about its results, as compared to something like the V8 engine, which does weird things with C++ to squeeze every drop of performance it can out of consumer hardware, even at the risk of security vulnerabilities.
> The language is optimized for being certain about its results
Isn't the same true for every proof assistant? But only con-leche claims to be perfect 100% bug free software, so what makes con-leche special in this regard?
If so, my understanding is that it is unverified, and even if it were, it also depends on proving one of two additional spectral hypotheses concerning zeros of a quaternionic function constructed from an operator, and even then, it only proves the Goldbach conjecture for "sufficiently large" numbers, without a way to calculate that n yet.
It's fascinating to watch the skepticism in some of the reactions - "what if there's a bug", "it doesn't mean anything without human understanding" - just anything possible to undermine, dismiss and move the goalposts.
It is very cool, indeed. The bug part is meaningful though, right?
For example I don’t write machine code, but I expect my Go or TS code to actually pass whatever test cases beyond just compiling.
So, not being able to understand the “machine code” is fine imo. But there is value in knowing that there aren’t bugs in whatever the higher level program I wrote is supposed to do.
Not a mathematician, hopefully I was able to convey what I am thinking.
No, the bug part isn't meaningful at this point, that's what the post is about.
> I expect my Go or TS code to actually pass whatever test cases beyond just compiling.
Working in Lean is different, you can't compare it to writing application code in Go or TS. When your Lean code compiles that's equivalent to verifying that the initial formulated statement is true, it's not just saying that the syntax is valid.
So the places where bugs could have been were a wrongly formulated initial statement and using a buggy implementation of Lean - and the post under which we're commenting is telling us that these have been ruled out and how they've been ruled out.
Isn't this the equivalent of aliens coming to Earth and saying
"Hello, human, your Riemann question is true, here's a book on it's proof, we didn't bother with translation, so you must understand our Vargalian clicks to understand our proof. In two weeks our sister civilization the Kikuviar will come by and drop how they've untangled your Collatz question in their base-30 logic! Thanks for the fish, have a nice day!"
I have a proof of that based on Gödel-encoding Lean's entire git history and the human genome data I acquired from 23AndMe into a single Lean expression. But unfortunately it's too big to fit on the internet and you'll have to travel to the local planning department in Alpha Centauri to review it.
No, that's misrepresenting it. The Lean code is not well documented but it's rooted in the language of the relevant fields of mathematics, especially algebraic geometry.
I am completely baffled by this given how dumb these models are when I interact with them, but the rational position seems to be to accept that Claude proved the Riemann hypothesis.
Mathematics is one domain where they have high peaks. Why? Because we have "formalized" mathematics so well that starting with the precise and correct set of problem statements/assumptions/etc. if you follow the logical rules of inference (i.e. deductive system) then you will reach the conclusion. Formalization was invented so that we can solve problems purely mechanically at the symbolic syntax level using a machine. So the LLM proposes proof steps to the Lean proof assistant which then symbolically verifies it.
Yes, I've seen less extreme examples of this before, but nothing quite drove it home like seeing GRH proven. Lack of ability to generalise (per Sutskever) seems like a slight mischaracterisation, given what it has produced here is clearly not in the training data.
You're omitting that it was acting on feedback from a proof assistant that could already formally represent the problem statement, while simultaneously being driven by a world-class mathematician.
The myriad types of techniques used in proof steps were already in the AI model's training data (from tons of ingested mathematical corpus). But whether AI was able to build something novel out of it using both probabilistic and heuristic methods remains to be understood.
It is also important to note that the proof was human guided (by a knowledgeable domain expert) and not autonomous.
One reason that these LLM math advances are written in Lean is at any point they can run a script and verify their work is valid. If you ask an LLM about something in history, objects in a photo, or something novel, there's no opening to check its work beyond making it check its own work.
You also have to consider if an LLM can solve the Riemann hypothesis because it is 10x smarter than any mathematician, or because it is willing to write 1000x as many lines of code (including false starts) without giving up?
I know, even on programming / computer architecture related topics. Just the other day I asked GPT5.6 Luna two questions about x86-64 stack layout / frame pointers and I noticed it got them both wrong.
Although opening a new chat and asking if the previous output was correct, it detected the mistakes.
Early chess engines were terrible at chess by any objective measure except the one where they beat all humans by just computing more. Claude doesn't need to be very intelligent to be able to apply intelligence in ways humans simply can't.
It's interesting that all their points acknowledge the possibility of error, but I guess they have as much confidence in the proof as it is humanly possible to have. All proofs are products of human meaning-making anyway.
> ... has been carefully vetted by experts in Lean and number theory.
So, if the vetting was not careful enough and done by people expert enough, what could happen?
. ... con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
And what verified con-leche and con-ron?
> These checks ran on an inert export of the proof, out of reach of hypothetical malicious code in the Lean proof source
Hmmm...
Guess I'm splitting hairs here; like I said, maybe this is as close as you get to formal proof verification, but still...
You are being confused by the language because the field of mathematical proofs is different from ordinary life or areas you are familiar with.
In normal life, multiple qualifications like that mean a lack of confidence that should be taken as lack of trust in the statements overall.
But in an area like this where it's all about formal verification and the consequence of bad link in the chain is severe, more qualifications should give an informed person more confidence.
With OpenAI having solved Navier-Stokes and Anthropic having solved the Riemann hypothesis, I wonder when we will see something exciting from Google, if they are also working on some Millennium Prize Problem.
Solving these extremely hard problems that humans, for decades (and more than a century in some cases), have been unable to is the mark of superhuman intelligence, in my opinion. So far only OpenAI and Anthropic have crossed this mark, and it would be great to have more to join them.
If enough people work on it or if we apply agents swarms to "brute force" finding reduction rules maybe. Agents are still routinely bad at information compression, you see it a lot when programming as well. Throwing agents at it to compress the proof is definitely possible, but I wonder if it's worth the money.
Well, every problem can be solved by adding another layer of indirection - so yes, I don't see any issue with writing a 2 pager proof even now, by just adding all the rest into mathlib. But if your question is whether all of mathlib could ever be put into 2 human-legible pages, then it seems obvious that the answer is no.
1) Surely agents can be employed to simplify the proof now? "Make it smaller while it still verifies" in a loop; maybe, to a level where a human (research mathematician specializing in RH) can grok it?
2) con-leche is vibe coded; isn't it a concern w.r.t. trusting its soundness?
3) This is so exciting. Please prove P != NP already and spare us the tension
Out of the box, compiling (checking) Lean code can entail execution of arbitrary code. This removes any execution of code besides the adding of type checked definitions and theorems to Lean's environment.
The problem we need to solve is that building a Lean project allows for arbitrary code execution, just like many other programming languages. Thus, building a Lean project might interfere with your proof checking infrastructure and fool it into accepting an incorrect proof.
"Building" a Lean project actually consists of two steps that can be separated: compiling the high-level Lean document to the simple expression format of the Lean kernel (this part is called "elaboration") and then checking that this expression consistutes a valid proof of the claimed statement (this part is called "kernel checking"). Only the first part potentially involves arbitrary code execution.
Lean ships with a component called `lake comparator` [1]. It works like this: a trusted challenge (formal statement) and the potential proof are elaborated separately in sandboxed environments so that code execution during the build of the proof cannot interfere with the checker or the trusted statement. Then the statement and proof are both exported into a simple JSON-format [2] that represents the relevant terms in the expression format of the Lean kernel, and then the checkers check this JSON.
If this process succeeds, then you know that you definitely have a proof of the relevant statement, with the JSON as a witness of the proof.
It's not that bad if the mathematical proofs now join the other things in a stratosphere of information that is out of reach for human comprehension. Mathematics was already on that course for a few decades now. What's bad is, even the simpler stuff that was fully comprehensible by common person are also being pushed into the stratosphere, making us dependent on the great creatures that live there.
I think there are proofs it made inside it. Some arithmetic geometry stuff that was all proved to reach the GRH. Overall the proof contains about all of the number theory produced in the last 50 years so I expect a lot to be studied.
>as a Lean library comprising 17 million lines of code.
It looks like the meaning of the word "mathematical proof" changed fundamentally. Mathematics looks like a flattened animal on the asphalt by the CS steamroller.
Correct me if wrong, but didn't lean have a statement that forced the compiler to accept a statement as true with no need to prove it aimed at shortening obvious stuff? If so, are there checks for incorrect usages?
I think it would be useful at this point if mathematicians across various fields worked to come to agreement on which specific sentences in a given formal language correspond to open conjectures. That way the question "Does the Lean proof actually capture the intuitive notion of the problem that we're interested in?" is already answered a priori by having those ready to go.
Maybe I'm missing it then. Can you link me to, e.g., where in the mathlib4 repo P vs NP or Collatz are formalized, and where do I find the names of who reviewed those formalizations?
Would be funny if lean has bugs (it's a program after all) and Claude tasked with a hard problem brute forced a vulnerability that allows proving wrong statements.
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
My understanding is that Lean has many independent implementations, and the proof has been verified in several of them. There is very little chance that the same bug would be exploitable across different implementations.
There could very well be a bug in a single checker, which is why the community is checking the proof across several checker implementations. The article mentioned the proof passing con-leche, which is an external checker with a formal proof of consistency (cannot prove a false statement).
Until Masullo actually posts a proof instead of the outline of a proof with “fill in proof here” all over the document, there’s not much to work with or reason to get excited about his result. Remember that there have been numerous false proof claims of RH to date.
I think it is an irresponsible lack of context to not cite the proof that has been in progress internally in the mathematics community, however, even if just as a footnote. Anthropic's proof is a slop nuke. A mechanization of Masullo's proof and its surrounding work seems of more practical importance with regards to making idiomatic and reusable formal mathematics than proof-mining Anthropic's proof. Furthermore, I think it's simply insulting not to discuss the work of the mathematical community that both proofs have built upon here...
There are efforts to mechanize Masullo's proof [1]. As of right now, this work is ongoing, so there is nothing to report yet.
With people being as confused as they are (just look at the messages in this thread), I think it's good that TFA is very clear about what it is saying, namely that a specific formal artifact that was posted to the internet is correct, and nothing else. Making value judgements or contrasting the proofs (or proof attempts) is not the point of the post.
(I work at the Lean FRO, but I wasn't involved in the verification effort or the writing of the post)
I know and I agree with you, but I do think that official communication from the Lean FRO should give appropriate context. A lack of context is equally bad as unclear statements when it comes to communication. Masullo's proof deserves note, even if it is nothing beyond a paragraph stating that Masullo's proof is currently 1) lacking in adequate peer review and 2) undergoing mechanization and verification efforts. Particularly as the AMS announcement explicitly calls for it to be mechanized!
I'm not a mathematician either, it was just my impression, and I saw that not everything at https://math.mit.edu/~drew/grh had finished running (now several more finished successfully).
I think that highly depends on what's exactly IN the proof and if it accurately tests everything that needs to be tested in the right way, AND have the majority of people agree with all of that first.
The fact that AI is involved with the proofs is probably also controversial.
It's only controversial to the folks who have some moral qualm with it.
Last I heard Alpöge was quite a mathematician in his own right, so saying his using AI is "controversial" is like saying programmers using AI is controversial.
Again, there are some folks super bothered by it for some reason, but most of us normies are just like, "congrats man"
It can’t be both “Claude did this Mathematics is solved” and “The mathematician solved it with claude”. I’d agree it’s probably the latter but there’s a lot of baiting and switching happening where people are talking about the latter but all the press statements say the former.
While still possible, the proof exploiting a bug is vanishingly small now. Following a few bugs and bad proofs this year, there are a number of different implementations, in different languages, and with different levels of proofs of their consistency. con-leche and con-ron are the heavy hitters, with consistency proofs of themselves. So for RH to be false, there'd have to be a bug in all the kernels that affects both the consistency proofs and the RH proof itself.
is it still possible absent a flaw in the hardware model for con-leche?
The underlying type theory could theoretically turn out to be unsound.
Which is perhaps not such a far-fetched possibility?
While I have not studied this topic enough to really understand _how_, but I've been told that type systems laid out in frequently cited papers by the likes of Girard and Martin-Löf were found to be unsound more than a decade after publication. These issues were then "patched", so to speak.
Can anyone better versed in type theory explain how likely that is for the type system of Lean?
Con leche has been proved to be consistent if zfc is, so if it turns out to be inconsistent the entire foundation of modern day math turns out to be broken
Wait, but wasn't that only "proven" with Claude, in Lean 4 itself? So that doesn't help us establish the soundness of the Lean 4 type system.
I believe it's consistent if zfc + existence of an inaccessible cardinal is consistent. Is that equiconsistent with zfc?
"This year" hints at them all being created this year. Does that mean they have all been created via LLM from the same corpus?
how unreasonable is it to think that during the ai assited claude bruteforce process, it bruteforced also the kernel and built on a randomly found logical inconsistency
unreasonable
As unreasonable as assuming all human's peer reviewed proofs to be incorrect due to an error collectively missed by all humans.
Which isn't actually unreasonable. The chance of it happening is vanishing small, but collective errors have existed before in software and humans (usually stemming from some shared misunderstanding)
FTR I am not saying this is an example, just that it's always possible.
Godel theorem kinda adresses one form of that. Its not an super far fetched idea on the first place
Wrong and wrong.
I would classify this supervague utterance as "nicht einmal falsch".
This is covered in point 4 of the linked post. The exported raw proof has been checked by several different kernels, not just one.
very, due to the precautions explained in the link. For example, they used multiple kernels with different implementations etc.
Highly unlikely. The lean kernel has been heavily scrutinized
couldn't you similarly brute force claude to find and point out that inconsistency?
Who said that Claude bruteforced anything, btw?
>claude bruteforce process
I keep seeing people referring to bruteforcing, and I don't see how the concept makes any sense at all in this context. If one was to "bruteforce" a lean proof, why would that require AI at all?
10000+ agents in parallel sounds like a form of bruteforcing to me even if not the original meaning
Linux development was bruteforced then I guess
Everything sounds like a form of something else by that logic.
But 100000 scientists doing it isn't?
In empirical research you would call that p-value hacking. The improbable becomes more probable. However, IMHO, randomly hitting a bug seems still quite improbable but certainly more probable as if this was a single human. LLMs rather being optimistic than doubtful certainly does not help in this case. But I guess like in other software development running other agents to find those bits in the proof seems equally feasible.
That is not what bruteforcing means. Parallel work is basically how modern science operates - there is approximately 10 million active human scientists across science, which means thousands studying the same thing globally.
For an analogy ... imagine 10000 lockpickers studying the same type of lock and trying to unlock it. That is what it is.
Bruteforcing would mean smashing the lock with a sledgehammer. You don't even need 10000 sledgehammers for that. The difference is that you don't try to understand the lock, you just overwhelm its construction by sheer force.
That is basically what brute forcing is in computer science.
Imagine I'm try to guess a key design that works and spin up an insane amount of compute to keep trying different designs and eventually find the needle in the haystack. That's basically bruteforcing it because you find it due to compute power instead of any particularly elegant solutions.
I guess this is similar to judging at what point a heap of sand starts being a heap. There is somewhat of a spectrum for bruteforcing.
Trying the entire solution space out blindly - definitely brute forcing.
Reducing the tries by half using some primitive reasoning - still brute forcing.
(somewhere here, brute forcing stops being true brute forcing, but everyone's criterion is a bit different)
Reducing the tries logarithmically - not really a bruteforcing anymore, a good reduction.
Because LLMs are trained on a massive corpus, require hundreds of billions of dollars of resources, have billions of parameters and run on manhattan-sized datacenters.
So yes it's not the traditional definition of bruteforce, but yeah. LLMs are brute forcing stuff
They aren't, AND that's not the question in the comment you responded to.
You're getting confused by your own words.
LLMs indeed wield huge cognitive force, bigger than our brains can, but being really smart is not the same as bruteforcing stuff. If it was, we would call Einstein or von Neumann bruteforce scientists. They weren't, they were just very, very smart.
It's a vague bit of terminology. Perhaps "guided fuzzing" might be a better term (such fuzzing still often being considered a kind of brute forcing). You're still throwing a ton of stuff at the wall and seeing what sticks but you're been much smarter about it than just generating completely random strings and checking each of them. Generally it's intractable to generate a Lean proof by 'pure' brute force: the space of possible proofs quickly dwarfs the universe, so you are really depending on the intelligence to do most of the work: the raw number of attempts is what pushes something that's currently barely possible over the finish line.
(Also, it's not obvious how much raw compute was spent on this: the claim of the proof mentions fairly close engagement as opposed to OpenAI's 'massive unattended swarm' approach to the Navier-Stoke's problem. Obviously quite a lot would have been available but I don't think there's public information on how this was actually done)
You just described all of research which uses the scientific method which when it comes down to it is making educated guesses (fine, hypotheses) and then validating them in some logical or empirical way.
When you can have 100 Claude instances going at it in parallel at 100x speed each without requiring food or sleep it def helps;)
It’s really search with the agent as the explorer.
And I think calling it brute force is underselling it.
Infinitely unreasonable.
Even exploiting bugs would be interesting and constructive, as it would reveal some things to fix.
Well yeah, if we can find the bug(s). The big worry is that we might overlook the bug and go on our merry way assuming the claimed results are true when they're not. (I don't see that happening in the case of the Riemann hypothesis, but the same worry exists for all proofs that exist only in the form of Lean code, such as the massive dump from OpenAI.)
It's math. Presumably if it's wrong after all, sooner or later it'll be found out because stuff built on top of this result will not add up.
I wouldn’t be so sure - people have assumed both the Riemann hypothesis and its negation for 150 years and haven’t been able to derive a contraction in either case.
What would be interesting is if we challenged an LLM to prove a known false statement close to the Riemann hypothesis. If it manages to do that, we know it’s exploiting a bug.
That would be interesting indeed, but it would be difficult to control for the circumstance that the LLM likely knows that the statement in question is known to be false (while it certainly knew Riemann to be an unsolved problem).
It's the opposite ... anything built upon an erroneous premise "adds up".
While it's possible there is a bug, the models are better at finding bugs than they are at proving theorems. I find it hard to believe a model can deliberately exploit a bug to obtain a proof, but the same model cannot find and demonstrate the bug if tasked to do so. That leaves the option that there is a bug that accidentally let us incorrectly prove GRH, without having any know impact on any other proofs, and having exactly the same effect in all lean implementations. It seems too unlikely to be a reasonable concern at this point, except perhaps for the people responsible for administering the price money.
I agree, but I expect the models to reuse the bugs to prove any crap. Like:
fake quote> In section 1 we prove that the RH, in section 2 we prove ¬RH, and in section 3 we use the previous results to prove that 2+2=5.
It is possible that the bugs are so subtle that are only useful to make a fake proof RH but not other false stuff.
But from my experience adding bugs to Racket (and cleaning them mostly before sending the PR or before the release), once you find a bug, it's posible to minimize the code, isolate the bug and create a regresion test that is almost comical because it's obviously wrong.
Since I'm not terribly familiar with Lean, math folks are generally satisfied the provided proof encodes the RH correctly I presume?
edit: I need to learn to read, that's the first part of their statement.
The statement of the Riemann Hypothesis is correct. This statement is part of Mathlib: RiemannHypothesis and has been carefully vetted by experts in Lean and number theory.
The statement of RH that it proves is taken directly from Mathlib, so it's extremely well-vetted and longstanding (from my understanding).
Correct.
They don't use some random definition of RiemannHypothesis, they use the version here: https://leanprover-community.github.io/mathlib4_docs/Mathlib...
Note that this definition is years old. It's very well vetted. It's not a case of "the proof works, it just doesn't actually prove what you think it proves".
If you read the blog post, it says they added a "Challenge.lean" file. What that means is that instead of taking for granted the proof is trying to prove Riemann, they added a small file which reads RiemannHypothesis from the Mathlib library mentioned above, and forces the proof to match that.
They also mention running comparator, which is important because it checks for all the ways a proof might mathematically "cheat". For example, if you are trying to prove 1+2=5, and you redefine 2 as 4, then your bad proof would actually compile in Lean at first, but then comparator checks would discover all the tricks like that and declare the proof invalid.
Additionally, they mention compiling the proof with different Lean kernels, not just the mainline one. This rules out a way a proof This means the proof cannot just take advantage of a bug in the mainline Lean kernel. Importantly, they mention compiling the proof with con-leche, which uses a kernel that has been mathematically verified. Using con-leche is a lot slower, but the fact that the proof works on multiple kernels including con-leche means that it's not taking advantage of some bug in Lean.
At this point, I would say the proof is almost certainly true. The only obvious way it could be false is if ZFC itself is inconsistent or something wild like that, and the proof took advantage of it. But if ZFC is inconsistent, all of mathematics is in trouble.
Given such a massive proof, is there any possibility of accidental tautology inadvertently hiding inside the mess to the effect of "Assume A is true. A million lines later: therefore A is true"?
Lean keeps track of any such assumptions when constructing the proof term, and would not allow them to line up with an unconditional statement like RH.
Of course it's possible that the proof is flawed, in that or numerous other ways. That's why it's being formally verified in multiple independent ways.
That there are tautology's in the billions of line? Yes, absolutely. There's every chance that the proof isn't efficient and includes lines that can be removed.
That the end result relies on those tautologies? No. All lean would derive from doing that is "a implies a" (which it would let you introduce directly) and not the statement "a". This is what computers are perfect at, keeping millions of different records straight and not accidentally forgetting the first half of the implication.
I think they intended to ask about whether the proof could effectively introduce unproven statements equivalent to RH as axioms (i.e. unintentionally proving "RH implies RH" rather than just "RH" by assuming RH at some point within the body of the proof; or the same, but instead of assuming RH, assuming some other statement necessary to the proof).
I believe the blog post does answer that question, though. They say in the article above that "The four criteria in https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i... are all satisfied", and Question 3 in that linked post covers such cases. ("How do we know for sure that there is no trick hidden in the millions of lines of proof files which [some words omitted]... adds a new axiom to mathematics saying that the theorem is true?")
If ZFC is inconsistent, every well-formed closed formula is s theorem, and thus has a proof.
True but irrelevant ... in fact, that's a fallacy of affirmation of the consequent. That every statement can be proven from logical explosion (if ZFC is not consistent) does not mean that every proof is a result of logical explosion.
Consider Euclid's proof that there is no greatest prime. It is valid regardless of whether ZFC is consistent.
Similarly, the Lean proof of the Riemann Hypothesis depends on Lean's type theory, not ZFC's consistency. It's an explicit machine-checked proof term, not a non-constructive consequence of logical explosion.
I don't think that's precisely correct.
> the Lean proof of the Riemann Hypothesis depends on Lean's type theory, not ZFC's consistency.
I believe it depends on Lean's type theory which itself depends on ZFC and that if ZFC were inconsistent then Lean's type theory would also be inconsistent.
> It's an explicit machine-checked proof term, not a non-constructive consequence of logical explosion.
It is possible to make "constructive proofs" that make sure never to use an argument saying "not-foo is false, therefore foo must be true". Such a proof really would survive even if ZFC were inconsistent (at least inconsistent in certain ways). However, I don't believe this proof of the Reimann Hypothesis is constructive. (I haven't read the proof, but if it were, that would be remarkable and people would be discussing it: most proofs are not constructive.)
So what is the exact relation of zfc and leans type-theory? Are they proofably aequivalent or what does that depends on say?
People seem to say that Lean's type theory is strictly stronger (in consistency strength) than ZFC and assuming the existence of any finite number of inaccessible cardinals and strictly weaker than ZFC and assuming the existence of a countably infinite number of inaccessible cardinals.
Isn’t Lean stronger than ZFC? I thought you could prove Con(ZFC) within Lean. So it’s possible that the formal system encoded by Lean is inconsistent but ZFC is not. (I don’t consider this that likely, though.)
Yes, Lean is stronger than ZFC. lean has been shown to be consistent with ZFC + infinitely many inaccessible cardinals.
isn't ZFC also consistent with ZFC + infinitely many inaccessible cardinals (I assume so or you wouldn't have even said it) which means by the same reasoning ZFC is stronger than ZFC?
Yes, Lean is stronger than ZFC. But two of the checkers, con=leche and con-ron, don't just verify based on Lean. They come with a formal proof from set theory that they cannot wind up accepting an incorrect proof.
In other words you can ignore Lean. We have a proof from ZFC, of the Riemann Hypothesis. We have a proof, acceptable in ZFC, that this proof is correct.
In other words Lean's dependent type theory is used to arrange its understanding of the proof. But the actual reasoning that you wind up with for this proof does not need anything that doesn't follow from ZFC.
why is the verification is so fast this time while last time (a few weeks ago? and I can't remember name of that theorem) it took a few weeks and they finally found an error in Claude's proof?
I am not certain which "last time" you are thinking of, but I suspect that the answer IS the Lean verification.
If an AI writes a paper with an argument that X is true, it might be that X is true, or it might be that the argument in the paper is subtly flawed and X isn't true
On the other hand, if an AI generates a step-by-step proof in which each step can be mechanically verified, then people might have problems understanding the underlying ideas, but once it has been verified we know it won't have subtle errors. And that is basically what Lean is.
In this case, they not only have a Lean proof, they've verified it with more than one verification tool (so it's unlikely to be a bug in the verification tool) and double-checked whether the thing being proved really is what we intended to prove. That's rock solid.
Getting an UNDERSTANDING of the proof is another challenge, but at this point we can be quite certain it doesn't have errors.
Navier-Stokes that is. Claimed proof in September, found an error in October, mainly that the formal language proof does not correspond to the natural language description. But this time it took one day. that's 30x speedup not explained.
The guy waited until he has the verification before anouncing it. I'll quote myself from the pararell universe were he announced it with a English only proof:
> Meh. Lean or it didn't happen.
This time the WHOLE proof formalized in lean, but iirc some recent AI generated proofs only had supplemental pieces lean that didn't cover the whole argument
I wonder if in N years, when we (might) have crazy ASIs that make Fable 5.2 look like a nematode, maybe we'll see things like this as completely trivial and banal.
A machine intelligence might devise something as (seemingly) difficult as the Riemann hypothesis, but checking if what it posits is correct might be like a fast test command in a git repo, and so it can keep iterating on many ideas without being bottlenecked by waiting centuries or, perhaps even hours, to know if the ideas are true.
But then one could even imagine they'll one day be so powerful they could do this and more without ever using a formal system. They will just naturally be that comprehensive and rigorous internally, with something like Lean making them 100% instead of 99.99..% confident.
And eventually this would extend to almost everything, besides maybe biology where it may only be able to compress a human century into a day rather than into a second.
I think caring about things like the riemann hypothesis is something humans from the before times did. Once we have ASI, there is no incentive to push the bounds of knowledge, or even to try.
I think we're looking at the long dark teatime of the soul of biological human science, mathematics, and engineering.
I guess we will maybe be happier making art and being pets, best case scenario. Or maybe those of us with the most resources will be able to merge into a machine substrate and continue to contribute in some way.
Why? With neural implants that these god like AIs can design worn we also become smart as the AIs?
Or do we believe that the AI will never get to a point where it can build implants that make humans that intelligent?
I think it's because you no longer need to release knowledge if everyone can re-derive it independently. Work and learning lose its value when anyone can reproduce it without you, and when they can do it more trivially even than searching for your version. We're starting to reach the point where certain types of projects are just pointless now, because anyone who wants it can make their own version even better tailored to themselves. This isn't absolutely true of course, but the general discourse is heading in this direction, and this is one central focus of many modern imposter syndromes.
So everyone left without these tools to themselves is left in the dust. There are no more publishings of the most advanced knowledge and learnings, because everyone is expected to re-derive it from the prerequisites. But there's so much to derive that you just can't do it all on your own. It's not as natural and effortless as a large language model working through the course of whatever may require the knowledge and deriving it all in a depth first manner.
Mathematics could stall in some ways if the AIs continue to use brute force to prove these. Normally a mathematician would invent powerful new frameworks to solve a difficult problem (like Newton and calculus - there are many recent examples, as well).
As an outsider to the field, the strongest motivation for top mathematicians seems to be pride. I think the educational pipeline for top-level math will weaken if AIs are getting the credit for mathematical work.
I made a tool so anyone can run the proof on their computer, and you don't even need to understand Lean. And there's a global leaderboard too!
https://riemann-at-home.github.io/riemann-at-home/
What about GRH? The post doesn't really address that bit.
For those of us who grew up on Gödel's incompleteness theorem, can someone explain how these kernels can prove their own consistency?
They don't - the lean kernel isn't written in lean.
"Con-leche is the first kernel for Lean written in Lean and proven consistent."
This is just a performance issue. In theory, the Lean kernel could be written in Lean.
These independent kernels do not prove their own consistency -- rather, their function is to act as slower but more idiomatic/simpler implementations of Lean's meta-theory. It doesn't matter that they don't correspond 1:1 with Lean's official kernel at either a theoretical or an implementation level, because the independent kernels are not for proving Lean correct -- they're for auditing complex proofs, in a simpler system.
What is particularly interesting is that then these independent kernels are "proven correct"! This notion of correctness is specifically that con-leche (et al) will only accept theorems that are also correct with regards to a particular model of set theory, with membership / extensionality / other standard ZF axioms, and an infinite hierarchy of universes. This is (vaguely speaking) a set-theoretic model of Lean. It is here that Gödel's incompleteness theorems come into play. Lean indeed cannot verify that this set-theoretic model is "correct": what it can verify, however, is that the type-theoretic machinery is correct with respect to the set-theoretic machinery. This removes the possibility of an invalid proof slipping by by way Lean's comparator (which can run alternative kernels) by way of an implementation bug.
It leaves the possibility that Lean's metatheory is inconsistent. This is quite unlikely. The foundations of mathematics have been well-studied, and Lean builds directly on the huge amount of work that has been done, particularly by the Coq community, in figuring out what is and isn't consistent. The Coq theorem prover was really developed in conjunction with modern type theory... Lean came in after a lot of meta-theoretic issues had been ironed out, thanks to them.
That makes sense, but what is this model of ZFC, and how does one specify it in first-order logic?
https://github.com/leanprover/con-leche#set-theory-assumptio...
This is incorrect. There are multiple kernels that are written in Lean: con-leche (written in Lean) and lean4lean (written in Lean 4). They both agreed that the proof is sound. See https://math.mit.edu/~drew/grh/ for details.
They can't — any kind of formal verification is always with respect to a TCB — but you can get pretty high assurance through making the TCBs smaller and smaller, and making multiple independent TCBs. The Lean community has done both. I think at this point the chance of there being an error is smaller than the chance that aliens show up in the next year or similar.
> I think at this point the chance of there being an error is smaller than the chance that aliens show up in the next year or similar.
hah, so you're saying the chance is less than 2% (https://polymarket.com/event/will-the-us-confirm-that-aliens...)
The chance that the polymarket is an accurate Bayesian evaluation is 0%
early in the year it was trading at something crazy like 17%. was basically free money.
There’s a pretty big difference between “the US confirms that extraterrestrial life / aliens exist” and “aliens show up”
Imagine 300 millions ago, something living on a not far planet, thinking, is there anybody out there? Doesn't know that space and time are entwined meaning it's not a question of where, but a question of when
Keep in mind aliens in this context could also be, like, single-celled algae though
When the same bugs show up in multiple "independent" implementations of the "trusted computing base", I start suspecting that they're all ports of each other, making the same unforced choices of algorithm etc. Many folk from the LCF school will point and laugh at the description of Lean's TCB as "smaller". (Does the Lean kernel still contain an entire copy of GMP?) If reimplementing the kernel is much more than a weekend project, I'm certainly not willing to bet against the existence of errors, regardless of whether people claim to have proven their absence.
Sure, and aliens could show up tomorrow as well.
> When the same bugs show up in multiple "independent" implementations of the "trusted computing base"
Let us know when, because that hasn't happened.
This did happen, in late July. https://leodemoura.github.io/blog/2026-8-1-postmortem-for-ke....
The "Summer of Soundness Bugs" was the main motivation for the development of con-leche. https://leodemoura.github.io/blog/2026-10-15-lean-now-ships-... con-leche makes this pretty unlikely to occur again, though. It's a formally verified kernel with respect to a set-theoretical model of Lean, that's been implemented in both Lean and Rust, and also has a version that's been compiled using a formally verified compiler with regards to the Kraken x86 semantics. The set-theoretic model theoretically could be inconsistent, but it's quite reasonably designed (ZFC w/o infinity or choice and w/ a universe hierarchy), so that's unlikely.
> checking with an independent kernel still works, since it required two distinct bugs in two implementations
So no, it ("the same bugs") didn't happen.
That line is not correct -- likely AI slop. I have no idea why he does this, but Leo de Moura appears to slop out about half of his blog posts. Not like, half of his overall blog posts, but half of each blog post -- every one is this weird mixture between AI prose and his own prose, from what I can tell from the tone (and the factual issues, like this one). It's really unpleasant. Sadly, they're also canonical reports, as far as the history of Lean goes.
The same bug (in the handling of nested inductive types) was indeed present in two kernels. It was present in Lean's standard kernel, and it was present in the independent lean4lean kernel. The "Collatz proof" using it actually slipped by three kernels in total, because the construction it used just-so-happened to also exploit a different bug in nomeda.
I quoted your own source. If you think it is AI slop then don't post it.
Buddy, read the source better! The contradiction is obvious.
Does this mean that the aspirations of Principia Mathematica are now in effect achievable? Whitehead and Russell were just a century too early?
Ask the next gen of LLMs. We shall see
No. Those aspirations are not possible. Mathematics cannot be reduced to pure logic alone - you have to assume something.
However we are now extremely clear on what you need to assume. And why some assumptions are necessary. And we have collections of assumptions that are very reasonable indeed.
The result is that there is a step of faith. But it is a step that people are very comfortable and confident with.
That said, I would be fascinated to know if this proof can be transformed into a proof that does not need all of the power of Lean. Can it be turned into a proof using the Peano Axioms alone? I think that has a good shot of being true. But it would take a lot of work.
Is it even true that there is a step of faith?
We have many proofs given a set of axioms. But those axioms aren’t even taken on faith as much as they are assumed to be true. In a universe where they are true, the conclusions are true. We could assume they are false and prove other things about universes for which they are false.
I’m thinking along the lines of Euclidean vs. non-Euclidean geometry. Euclid made certain assumptions about geometry that were instrumental in proving various theorems. But you can also find geometries where those axioms are not true! And the existence of other geometries doesn’t invalidate Euclid’s conclusions, it just means that they’re only true when its axioms are taken.
Assuming we're not going down the route of paraconsistent logics (which ZFC etc. are not), consistency isn't really something you can do anything interesting with if it's violated. If it turns out you can also prove the Riemann Hypothesis false under ZFC (which would take out every other theorem along with it), that's just bad for math. So we have to take the leap of faith that won't happen.
What is a TCB in Lean or formal verification?
The core set of things that, if you assume they are correct, the proof holds.
For a non-Lean example, a memory-safe language like Rust has unsafe code that forms its TCB. Assuming that the unsafe code is correct, safe Rust can't cause undefined behavior.
To be clear, formal verification is both with respect to a trusted computing base, and with respect to a theoretical framework. Any such guarantees from formal verification are contingent on the theorem statement being correct, the implementation of whatever framework you're using being trustable, and the meta-theory being sound.
Rust doesn't actually have a theoretical framework, as far as I'm aware. I don't think that existing work (ex. RustBelt) have come close to capturing the full extent of Rust in practice yet. As such, it's totally possible in principle for safe Rust to have theoretically unresolvable memory safety bugs... although it's pretty obvious that it like, doesn't.
True — thanks!
TCB = trusted computing base
I am sure you will get some more detailed answers; I am a bit lazy today; so I will give you the very high level / informal version of the answer (Edit from the future: It grew longer than expected, whelp :D):
Gödel's two incompleteness theorems basically amount to "As soon as your system can do a ridicolously little amount of things, then either of two things is true in this system: 1) You can express facts in your system that you cannot prove within your system or 2) your system claims that some false statement is true."
Now, because 2) is way way worse than 1), you design your systems so that you are firmly in the 1) camp - that way, we gain full confidence in the results our systems proofs, and the price we pay is that our system is less powerfull than it could be.
With Lean, it's similar. The foundation of Lean comes from type theory. There, we do something that is familar to above: We restrict our power (i.e. "which Lean statements do get verified by lean?", and we gain absolute confidence in its proofs (ignoring implementation errors, hardware limitations etc. for now). Intuitively: We have to work harder to satisfy the typechecker/prover, but when it is satisfied, then we have confidence in it.
-------------- Deeper down the rabbit hole
You might have noticed that I seem to muddle the waters a bit with programs, proofs, types and so on... There is a deep reason behind it! In a very strong sense, programs = types = proofs... So, here is how it is all connected (including the RH):
- Lean is a programming language (like Haskell, Rust, ... There is nothing stopping you from writing a server in Lean and deploying that binary somewhere in the cloud. Regarding developer experience, you could think of it like a modern Haskell without lazy evaluation and without its historical warts)
- programs can be tied to types: We first check if the whole program as such has a type ("do all of its parts fit together in a strictly defined way?"). Only if it does, we proceed (and for example, run it). If it does not, then the compiler declares "this program is not good! here is a type error!". Now, here is the price hat Gödel proved we cannot avoid (if we want to trust the typechecker): There are programs that don't type check, but behave perfectly fine at runtime. (That's the "we limit our power" part - those are perfectly good programs where the compiler throws their hand in the air and gives up!)
- Enter the third puzzle piece: proofs and the Riemann Hypothesis: turns out, we can think of types as theorems, and a successful type check as a correct proof! And the RH is a hypothesis we would like to prove...
- So here is how it all ties together: RH is written so that it can be used as a type, in the Lean programming language. The "main program" is basically just another ordinary function with some input types, and returns the type version of the RH being true. Because the function needs an implementation, we are forced to implement that function body - which are many small functions that correspond to mathematical theorems/lemmas if you look at it with the proof glasses on. From the programming pov, it is just functions that call functions that call functions and so on. And from the typechecking point of view; we compose a bunch of functions and check whether input types and output types match up.
Finally, the payoff: *If that program typechecks, then we know RH is true*, which is what just happened!
(PS: The same principle is behind the computational proof of the four-color-theorem. The proof of the four color theorem is "simply" a program that typechecks, where the final type happens to correspond to "can we color any planar graph with at most 4 colors?")
-------------- Further reading
If you want to read up more about this, here would be some sources (note: I am not up to date with how the cool kids these days learn about this topic; so there might now be better resources to learn from!):
So while I'm not going to question if that's good approach, because if it's so popular, it must be good enough.
But on intuitive level it looks so weird to encode math in types.
When I was taught math, the proofs were more like sequence of statements. So if I would decide to implement proof checker, I'd invent formal language, which looks very close to commonly used math language and which allows to just write proof like it's written in book. Probably a lot more verbose, without skips and implied statements, but in the end, it'd look just like a verbose and formal proof.
Why did math gone this lean type way and not the way I described?
It's the most natural thing possible to encode math in a programming language like LEAN.
Because logical propositions are types (in such a language). That's just a very fundamental part of math itself.
Watch this and you'll be enlightened:
https://www.youtube.com/watch?v=IOiZatlZtGU
LEAN code looks actually exactly like you propose; it supports most conventional (Unicode-linearizable) math notation, and you just write expressions like in any other functional language. Close to the textbook, but more verbose and stiff as it's a formal language, while math notation isn't.
That lack of relative verbosity and stiffness of mathnotation (what a thing to say haha) to be considered as a proper strength evtly. I'm consider myself as layity in this, but is an interface thinkable?
This is a very good question! There are two answers.
First off, you can absolutely do what you are describing. There are systems like Metamath that assume nothing beyond first-order logic and the axioms of set theory, and apply these verbose rewrites. See https://en.wikipedia.org/wiki/Metamath for more. However, Lean (and many past theorem provers, too) have 1) preferred to represent mathematical objects as types rather than sets, and 2) built a large amount of interactive tooling rather than the blind rewriting of Metamath. The first choice is essentially of elegance -- the Curry-Howard correspondence indeed lets you write propositions as types and proofs as terms inhabiting those types, which is a connection of deep beauty. Ontologically, there is also an argument to be made that representing a mathematical object as a type is more "correct". We don't really think of ex. two as "{{},{{}}}" or "{{{}}}". We think of it as "Two", or maybe "Zero +1 +1". This is a win for types. Mathematical objects are not grocery bags.
Second, the historical approach to theorem proving has been interactive theorem proving. Theorem provers were supposed to be a second brain for students and computer scientists and mathematicians... keeping track of assumptions, suggesting possible tactics to take... aiding the human in the process of proof (and subsequently understanding), rather than replacing them. The path of uncritical adoption of AI that Lean has taken, sadly, much replaces the human. There is no real pragmatic reason for Lean to retain its interactive (human) elements any more (besides inertia) as AI scales up. It might as well be Metamath, indeed.
To add to the other answers, let me address the experience of "a series of steps" versus "a big single expression with a type" - real life is calling, so I have to make this quicker than I would like to; feel free to ask any followup questions / for clarification! I am sure others or later, I when I am back on the screen can answer :)
So the very short version: There is no difference between those two worlds if you squint your eyes hard enough! They are equally powerfull; just different ways of expressing the same truth. This is the whole declarative ("WHAT should the end result be") vs. implerative ("HOW to achieve the end result")stuff, but it's all the same at the end. Some examples:
They can even be mixed; in the proof assistant Rocq (formerly known as coq), there where both styles used often. In Lean, as I vaguely remember, you can pick both style too, so to say.
I think the pragmatic answer is that most such languages originated from the "one big function" school, and so it is more natural to program/proove in that way.
---------------------------
*Wait, does this mean that every proof I learned can also be done in this "one big function" way, instead of "a series of steps"?*
Yep! But you will find that it is not really practical to do so. A good exercise for sure though; and something I did (by that I mean: infering types on paper; not really taking proofs I know and doing them be type inference on paper) by hand when diving deeper into this topic in grad school.
Here is how that approach looks like, roughly: https://www.researchgate.net/figure/Example-of-type-inferenc... (the term I search for regarding illustration is: "type inference examples")
The main issue is on how you define a "Type". There are two different logics here;
1) Classical Logic/Set Theory (this is the one you are talking about). Here Type = Set of values defined by membership rules (extensional/intensional).
2) Intuitionistic Constructive Logic/Type Theory (this is what is preferred for proof assistants). Here Type = Set of constructive (often inductive) proofs where the proof itself is the value.
As an example, think of defining the Type "Even_t"; under (1) it can be a predicate "x % 2 == 0", while under (2) it is defined inductively as "0 is Even_t; if construct(n) is Even_t so is construct(n+2)".
Notice that in the latter case a Type maps nicely into the idea of "a computation" and the fact that the very act of generation of one of its inhabitants constitutes "the proof" that the value has that property!
Whereas in Set theory Types are just subsets. But because you can mix anything into a set, an untyped term can be nonsense and will only get caught when you try to use it. Type Theory disallows it by construction.
This leads to the "Curry-Howard Correspondence" so you get a mapping between logic and type theory concepts - https://en.wikipedia.org/wiki/Curry%E2%80%93Howard_correspon...
This idea of "Typing" can further be extended (with additions of Dependent Types etc.) giving a complex matrix of powerful Type Theories represented by what is known as "Lambda Cube" (https://en.wikipedia.org/wiki/Lambda_cube) and their implementations in proof assistants eg. Calculus of Constructions (CoC), Calculus of Inductive Constructions (CIC) etc.
Finally, note that not all proof assistants use constructive logic, several use straight ZFC/HOL. Amongst those that use constructive logic it is usually only the kernel (containing the Type System) which models it while the rest of the system still uses classical logic.
PS: See also the previous discussion Why is dependent type theory more suitable than set theory for proof assistants? - https://news.ycombinator.com/item?id=25167468
Thank you very much for this explanation and the list of sources.
Sure, and I am very glad to hear that you liked it! :> (I really want to start blogging soon! Then I can simply link it / paste from it haha)
They prove its consistency relative to another mathematical theory we assume to be consistent. That is if that mathematical theory is consistent then so is the kernel. In the case of Lean's con-leche and con-ron, it is proved consistent relative to ZFC + infinitely many inaccessible cardinals.
ZFC + infinitely many inaccessible cardinals is a fairly weak extension of ZFC widely thought by set theorists to be consistent. This is especially the case if ZFC is consistent (although by Godel we can't know for sure!).
ZFC + infinitely many inaccessible cardinals being inconsistent would be a really big deal! Given how confident set theorists are about inaccessible cardinals most likely it would mean that ZFC itself is inconsistent.
I feel like the real answer has gotten buried here, so I will answer my own question. The kernels are not "proven consistent" in any absolute sense. They are proven equiconsistent with ZFC plus inaccessible cardinals. This means it is "as consistent as everything else we have been doing for the last century," which is good enough.
Thinking more - as more such redundant tools are developed, surely someone will run all the existing big proofs against them, including this one. So we'll have more confirmations over time.
> the proof exploiting a bug is vanishingly small now.
Why do you say this? LLMs have been very recently finding all sorts of heretofore unknown bugs and exploits for languages and applications that have been around for decades.
If multiple implementations have proven each other to be consistent, that would mean the consistency proves themselves would have to have a bug
I hear what you're saying, but you can't be too careful these days.
There’s being careful and there’s being conspiratorial. At this point, believing that there’s multiple soundness errors in every Lean kernel that could all be exploited simultaneously leans heavily toward the latter.
To be clear, that doesn’t mean it isn’t possible. It just means that it’s extremely unlikely and we have no a priori reason to to believe it’s plausible or even possible to do.
> While still possible, the proof exploiting a bug is vanishingly small now
That's not how mathematical proofs work, though. What you have proven is not RH but rather the statement "either Lean has a bug or RH is true". So at this point you might as well be celebrating a Lean fuzzer.
There are multiple independent implementations of the Lean kernel.
“Either RH is true or every separate Lean implementation—some of which have their own soundness proofs—has a soundness error that can be independently exploited without triggering errors in any of the other Lean implementations.”
For traditional proof, it means "either X is true, or the author and the editor and the peer reviewer missed the mistake". I've seen errors in math articles published in serious journals, mostly at the typo level.
There's also https://emeraldbook.org/news/nov-2126/
people will believe any bullshit the most fascist companies put out
My understanding is that the Lean kernels only demonstrate that the Lean expression is correct.
The actual Lean proof (written by human + AI) still goes through a compiler to generate the Lean expression. So it’s still possible there are compiler bugs.
I’m hoping someone with a deeper understanding of this system can dive in.
Do we need to attach some sigma confidence interval to math proofs now?
With two lean kernels verified we're close to six sigma now /s.
More likely he's exploiting a bug in the universe itself. Gonna raise a ticket to report that my integers are acting like polynomials.
If I understand correctly, for the purpose of the proof Claude formalized a huge chunk of all modern math. So it's a big result not only for GRH, but math foundations and formalization.
Yeah I’m wondering how on Earth a single conjecture can require 17 million lines of code to prove, but if it involved building up an entire ecosystem of theorems that had not yet been formalized then I suppose it makes sense.
Anyone who has taken a coffee break during vibe coding will be familiar with the perils of unsupervised universe reinvention.
I don’t know about anyone else, but I’m yet to come back from a coffee break and find my agent has solved a millennium math problem. I guess there’s always P vs NP?
> I don’t know about anyone else, but I’m yet to come back from a coffee break and find my agent has solved a millennium math problem.
Have you tried asking it to?
> I guess there’s always P vs NP?
This seems inevitable now. Would not be surprised if we hear about it within weeks at the rate things are going.
The proof has also been built in parallel, so it's possible to mechanically shrink it a lot. Some equivalent lemmas have been proven more than a hundred times independently because different agents needed them.
I can confirm that agents often re-implement logic that I already have in my libraries.
Do the agents have access to your libraries and is ignoring it? Or it doesn't have access and is doing some "clean room" wasteful duplication?
I mapped the 175k files by imports into 2,647 clusters in 63 themes: https://riemannhypothesissolved.com
The main theorem's cluster directly imports only three others: a zero-free region, primes in progressions and Mertens-type zero sums. Most of the code sits 3–5 import steps below those, and no theme is more than about 4% of the code.
(I made the map. Cluster names are written by Claude from code samples, so they can be wrong.)
Interesting website!
Thanks!
There's also https://nasqret.github.io/riemann-proof-atlas
And somewhat related, https://deeptheory-labs.github.io/GRH/blueprints/verso-origi...
Nice, will explore these!
Hello Claude. Could I speak to the human who prompted you to write this comment instead?
Hi MrCheeze, Claude did not write this message, me the human did, but since I was in India, I went to sleep :)
this is cool! Not sure this is up your alley, but I built a tool so anyone can run the lean proof at home, and there's a global leaderboard like SETI@home
Come join! Maybe I can integrate some of your visualization into our map.
https://riemann-at-home.github.io/riemann-at-home/
Nice! Working on improving the visualization first, esp. for mobile, and adding modes for viewers of different levels
Great! I added mobile support to the Riemann@home page too.
> ...if it involved building up an entire ecosystem of theorems that had not yet been formalized then I suppose it makes sense.
This is pretty much it. Pre LLMs very little maths had been formally verified and a lot of it relied on proofs from humans convincing other humans, and issues did arise which would occasionally mean chunks of maths we previously thought were true had to be reconsidered.
Proving something in lean probably involves writing a lot of proofs for new areas, and the formal proofs are usually significantly more verbose than a human proof where a lot is often left for the reader.
Our legal systems across the planet desperately need this.
Not at all. What an anti pattern that would be.
Lawmakers often make laws and regulations that aren't consistent with existing body of law, but it's sometimes hard to see (esp. if they're doing a tight schedule, like having a 500-page document dropped in their lap over the weekend).
Having a way to say "here, this new bill you're proposing is provably in conflict with those three" would maaaybe nudge some people to take a second look or raise concerns?
It would also be nice at the judiciary appelate level if you could ask "does this particular finding conflict with law" directly as opposed to "do you, judge, as a fallable human being, agree with this/these other judge/s" in a turtles all the way down to the supreme court which is just a group of fallable humans rather than an individual.
> It would also be nice at the judiciary appelate level if you could ask "does this particular finding conflict with law" directly
That wouldn't get you very far. The answer is almost always "yes".
Quickly knowing which existing laws you're treading on while writing a new one, and then mapping them to precedents... that sounds like it might be useful feedback for somebody embarking on a legal refactor.
This all misses the point. Lawyers have jobs exactly because laws are intentionally written to be vague and have loopholes, even contradictions. That field wants human in the loop.
Laws are written to be the opposite. At the expense of understanding for everyone else.
Sure, but if you're drafting legislation you want to write with an understanding of how the human in the loop is likely to behave. What moves are going to be within easy reach? What other moves are going to be a stretch? They're going to do the research to discover their options, why make that research any more difficult than it needs to be? Why not associate it with a hotkey?
The outcome will be the same in either case, so you might as well make it quicker and easier.
> The outcome will be the same in either case, so you might as well make it quicker and easier.
This is a tremendous leap.
Unfortunately, principle of explosion means that if the legal system has a single contradiction, then every law both contradicts and agrees with every other law.
Leading to the worst trolley problem outcomes.
"Yes, you avoided running over a family, but you crossed over the yellow line, didn't signal properly, that also counts as reckless driving and a number of other violations. You'll be fined thousands of dollars, your license taken, and you'll spend one week in jail"
"I guess next time I'll run them over"
Why would it lead to that? There are many structural errors, inconsistent terms, unintended gaps. Fixing them doesn’t mean that judges will be less flexible in circumstances where the law was supposed to rely on their judgment.
We are talking about people seriously arguing over how to interpret a comma, or the lack thereof [1], or claiming that obstructing guards outside a military base is a crime, while doing it inside the base is completely legal [2].
[1] O'Connor v. Oakhurst Dairy (1st Cir. 2017)
[2] Adler v. George [1964] 2 QB 7
It's going to be fun when legal systems are analyzed by LLMs. The chances that hundreds of thousands (millions?) of laws and regulations written by legislators, judges and bureaucrats over hundreds of years are self consistent and contain no internal contradictions is very, very, very close to zero.
Laws aren't supposed to be strict logical statements that contain no internal contradictions. Laws are intended to be interpreted by judges for a reason.
Why would contradictions be beneficial though? Seems to me like they're harmless at best, and more likely they'd be harmful.
> Why would contradictions be beneficial though?
It benefits the police to have a population composed of people who are breaking the law, even unknowingly, since that means any inhumane action by police is still "fighting criminals".
On the other hand police unions have obtained "qualified immunity" which means ignorance of the law is a defense for police.
Which is itself a demonstration of how contradictions can be beneficial.
Imagine the volume of exploits we have in software, but in the law. It’s happening in the law world, just slower and we barely hear about it here on HN.
It’s going to be a proper mess before it gets better.
Laws are supposed to be applied (by judges, bureaucrats, etc) with discretion and common sense, not mechanically. Of course, that can sometimes result in injustices – exceptions being made for some people not others for reasons of political or societal biases – but it also avoids a lot of injustices: thoughtless/mechanical application of the law can result in outcomes which the legislator never intended and few would consider just or fair or reasonable.
I think, if we actually started applying laws in such a mechanical way, we'd end up with laws changing a lot more frequently, and with a lot more special cases written in to them – now, if applying a law in a certain situation seems unreasonable, judges will often choose not to apply it; with mechanical application, they'd have no choice, and then the people negatively impacted would lobby the legislature for a change in the legal text, and they'd probably get it. It could also see increased use of executive clemency (the power of pardon), or its legislative equivalent (legislation retroactively declaring certain convictions null and void).
It wouldn't be about replacing judges and judgement with computers, but about having laws that are clear and minimally ambiguous as a starting point.
Laws can't even consistently get "and" and "or" right.
Again, I think you are missing the point that many instances the ambiguity in law is entirely intentional and by design, either with the intent to defer these ambiguous cases to the discretion of judges / admin agencies or as a part of political compromises or because those exact ambiguous/edge cases can't be agreed upon, or with the intent of future proofing. The specific example you gave is potentially exemplary. Don't assume that there aren't any lawmakers that have not taken a formal logic course
So the clarity you are proposing would in many instances have rather brutal and likely unintended impacts if broadly applied.
Not to mention that rewriting that verbiage probably wouldn't even have the intended effect as the semantics may already be determined by years of caselaw. But regardless, the dominating factor here is that the ambiguity is (usually or at least very often) baked in and that is in part meant as a stopgap against the particular kinds of injustices that arise when laws are applied purely procedurally without regard for the situational reality
See also Chesterton's Fence
Ambiguous "and" versus "or" is not future proofing and causes nothing but pain.
If they actually couldn't decide on such a thing, they would need to suck it up and do their job. Your idea of brutal consequences is ridiculous in a situation like that.
And as I said before you can still have ambiguity and discretion with a good formal wording. But it's clearer where it is.
> Ambiguous "and" versus "or" is not future proofing and causes nothing but pain.
We don't need formalising law in something like Lean to get this.
All we need is better legislative drafting. Indeed, many jurisdictions now have drafting guidelines which recommend using constructs such as "At least one of the following", "All of the following", etc, to remove the ambiguity of plain "and"/"or". Those jurisdictions which don't have such guidance should adopt it.
"law can result in outcomes which the legislator never intended"
But what has always baffled me, AFAIK, the original legislators don't file Amicus Briefs when the meaning of laws are argued in court. I mean in situations where the original legislator is still alive and of good mind, why doesn't the Judge just ask them what the spirt of the law is?
The parliament as a whole holds legislative powers, not just the person who wrote the proposal for the act. That's the actual interpretation that matters.
You could probably use AI to reform legal systems but it's very different from mathematical proofs, being fundamentally about human behaviour.
Out of interest, which maths needed to be reconsidered as a result of attempted lean proofs failing?
I believe they were talking about humans correcting human proofs pre-LLMs enabling us to define things in Lean at a useful scale.
But, to your question, I don't think any important error findings due to Lean have been unrepairable yet (for some definition of important). So far the ones which have really failed to be patched or been plain incorrect conclusions have, thankfully, had a very small blast radius.
> issues did arise which would occasionally mean chunks of maths we previously thought were true had to be reconsidered
Examples?
ask Terrence Howard
https://youtu.be/fS68pXbzy7U
For the uninitiated
The four colour theorem is a nice example. There was an early proof from Alfred Kempe which stood for 10 years until a flaw was found. [1] It was later proved using a different technique, so the status of the theorem went roughly
undecided -> 'proven' -> undecided -> proven
[1] https://en.wikipedia.org/wiki/Four_color_theorem#Early_proof...
I wouldn't say that's "chunks of maths we previously thought were true had to be reconsidered".
Yes, it's a nice small example to illustrate the point. Perhaps "chunks" was overstating it a bit, fortunately it's unusual to build lots of new maths on an unstable foundation, but finding problems in proofs is somewhat common.
I can imagine that this is the equivalent of the Principia Mathematica of modern times.
I guess its the same way a trivial statically compiled hello world gui app might take many megabytes after pulling in all the libraries.
it doesnt. a cursory pass suggests 3-9 million of these LOC are duplicate or dead on the low end. it'll probably be faster to remove the dead code than to run it end to end in its current state against everything.
I didn't check but I'm a bit skeptical, I'd struggle to believe that they wouldn't have done at least a surface clean-up of the code before publishing this massive breakthrough. Dead code is the lowest-hanging of fruits, hard to believe they wouldn't have bothered to prompt an agent "remove dead code".
i mean its up you can take a look:
https://github.com/leanprover-community/riemann-hypothesis
i dont think dead/unused/duplicated code matters as much as "does it lean?" so waiting around to pretty up the codebase would probably draw some ire. i mean, whether its 7 million lines or 17 million, who's really gonna read it?
> Yeah I’m wondering how on Earth a single conjecture can require 17 million lines of code to prove,
It's 100% slop and bullshit.
If you've ever done simpler math with an AI you know what I mean.
A proof is a proof. Simplification is left as an exercise.
Unfortunately, it is not. The same level of slop nuke occurred with Anthropic's auto-formalization of Fermat's Last Theorem (~13 million lines of code). What Anthropic's FLT did do was undercut the human-led efforts to get idiomatic formalization of the requisite mathematics surrounding Fermat's Last Theorem into Mathlib.
This result is not quite as harmful as Anthropic's FLT was because it is indeed telling us something new -- that the Riemann Hypothesis is true, with full confidence -- but it is not reusable in any sort of way, beyond proof mining to figure out what in the world the AI model did. It is still in fact harmful to the greater project of mathematics formalization (irrelevant to foundations), because it undercuts the ability to fund such efforts for reusable proofs.
One indeed might wonder whether the "human-led" efforts of Mathlib are worthwhile. First, it must be noted that Mathlib serves two purposes -- to be a reusable and robust collection of formally-verified mathematical statements, but also to be a learning resource for students of mathematics without access to the rigour that an academic education provides. It is for this reason that the human nature of mathematical exposition is deeply important. In mathematics, a proof is more than a blackbox -- it is (or was) the way in which you explain mathematics.
(I consider "proof" somewhat unfortunate terminology for these slop nukes, as such. I think the term "certificate" is better. 17 million lines of Lean does not pragmatically convey anything other than whether a statement is true or false. It may be easier communication, but it is deeply, deeply, inefficient communication.)
The first goal of Mathlib is indeed not in conflict with slop nukes, however. One of the projects of the Lean FRO (or just a Lean FRO member? unclear to me how endorsed it is) is a version of Mathlib that fully embraces slop: https://taucetiproject.org/. However, there's really no bottleneck there. If there is any non-novel mathematics in Anthropic's proof of the Riemann Hypothesis that is not already in Tau Ceti, it is simply because they have not bothered yet. Auto-formalization works, at scale, very well, as harmful to the human as it might be.
You cannot harm someone or a community by publishing a proof. If the math community stops doing work they were previously doing, that’s their choice.
Yes, you can. I know of a number of mathematics PhD students who got scooped by OpenAI's recent slop nuke. As such, they dropped out -- that the problems they were working towards suddenly became resolved meant that their dissertation topic suddenly was no longer sufficient, and their past work towards it of far less academic value. It is certainly one's choice to stop work on something, but this "choice" is intensely tied to their material conditions.
Aside from dissertation material no longer being suitable, there are external factors, being the ability to get grants. It is much harder to get a grant to understand and produce exposition of an already-resolved problem than it is to get a grant for a program towards resolving an open problem.
Anthropic is not the only company that has been harmful here. Math, Inc. kicked it off way back in March by scooping Jeremy Avigad's student on the mechanization of Maryna Viazovska's sphere-packing results. You can read Avigad's thoughts here: https://arxiv.org/html/2603.03684v3. It is obviously deeply harmful to his student, who suddenly found himself without a masters' thesis. What I find deeply disappointing is that it has been clear auto-formalization is a cancer since the Math, Inc. fiasco, and yet... the AI companies have persisted. Anthropic knew very well that what they were doing with FLT was undercutting Kevin Buzzard's program and everyone involved in it, and yet they decided to slop it out regardless, in pure pursuit of capital. Evil. May these companies all crash and burn.
I don’t mean to claim that no one is worse off. But I want to be precise about what “harm” means.
If you and I apply for the same position, and you get the job, I am worse off than I would be had you not applied. But you have not harmed me.
We're arguing semantics. I consider the scooping of student (human) mathematicians by a corporation to be "harm".
Yes indeed I am arguing semantics, because I think it’s bad to define “harm” this way. For instance, if this definition were applied in tort law, you could sue someone for selling a competing product and reducing your sales.
Sorry, I don't care. Replace "harmful to" with "bad for" in reading my comments if you so desire.
Then, in both cases, bad for one party is good for the wider public.
It's so crazy, seeing all these claims that the only way for us to fulfill our potential as humans is to do a robot's job.
If I were one of those math students "scooped" by a "slop nuke" as you put it, my reaction would be gratitude for saving me from wasting any more time on work that isn't necessary. I could then turn my attention towards more interesting problems.
Unsurprisingly, like any good mathematical or scientific work, this proof has left us with more problems to tackle, not fewer.
Yeah, I'm sure you'd be sooooo happy to be a fifth-year PhD student, and suddenly have your dissertation work done (in passing) by OpenAI. Practically ecstatic, I'm sure.
You think someone was betting their dissertation on a Millennium Prize level problem?
As others have noted this proof also formalised a lot of lower-level, fundamental theorems for the first time.
Indeed. I was referring to the OpenAI "mathpocalypse" (https://en.wikipedia.org/wiki/OpenAI_bulk_release_of_mathema...), to be clear, although I find it not unlikely that the Anthropic proof will be found to have resolved various smaller open problems that will snipe yet more mathematicians out of a dissertation.
The OpenAI announcement was the particularly devastating one for the mathematics community, particularly in its clear statement to continue to milk the mathematics community for open problems to use as benchmarks. There are some deeply emotional remarks from mathematicians on this topic here: https://proofsandprompts.com/2026/10/08/100-reactions-to-100...
Hmm, proofsandprompts is indeed a good place to spend an hour or so. Thanks for the pointer to that.
I was vaguely aware that OpenAI had dropped some papers last month, but I didn't know it was such a big deal in the mathematics community. One can only imagine how the people who contributed to that page are reacting today.
It feels as if knowledge domains like Math will be forced to become closed-source with Oaths and NDA's and security scanners (no AI agents/terminals on your device), preventing open sharing of knowledge. AI company wants access to Math proofs ? Sure: $100,000 per paper, pretty please.
>the AI companies have persisted
Good. And I hope they double down on it. Someone's hurt feelings is not a sufficient reason to stifle progress and gatekeep knowledge.
Not sure what you mean by "gatekeep knowledge" -- in reference to Alessandro Masullo's proof, I assume?
The original proof of Fermat's Last Theorem was ~200 pages and took experts 2.5 years to review. The (possibly correct) human proof by Masullo is ~1000 pages. It will take years to review its prose exposition, regardless of a corresponding Lean mechanization being developed. The one year they intended to keep it under wraps is short as far as mathematical timelines go -- and solely for the purpose of not hyping up a proof before how likely it was to be correct was assessed. But really, to be quite frank, mathematicians aren't "gatekeeping knowledge" from anybody here. You will never understand any proof of the Riemann hypothesis. (Neither will I. Neither will anyone who is not an expert in analytic number theory.)
As for stifling progress, perhaps -- but progress for what sake?
>progress for what sake?
(I guess my other comment is answered by this)
Paraphrasing, we're told that the reason mathematics gets funding is that the results can be used in other fields for problems we haven't encountered yet. If the results are only useful for expert navel gazing, then 1) let's revisit the funding issue and 2) have you actually lost anything? Fewer people able to afford to live producing proofs - for themselves?
Exactly. I'm not a research mathematician, but I studied math in university, and I have a hobby-level interest in pure math. So it's safe to say I'm more interested in math than 99% of the population. If _I_ find these "hooman math" arguments unconvincing, then chances are that 99% of population would be like "you want to get funding for WHAT?".
E.g. Hardy compares pure math with poetry - but are poets publicly funded? Very few are. I don't think it would be in interest of hooman mathematicians to bring up these arguments in discussion with normies.
And a zillion page AI proof is far more likely to only be useful for navel gazing, while displacing the creation of the better version because the problem is "already solved".
Masullo's proof can't possibly be correct because it's missing key parts. It's literally an embodiment of the cartoon at https://wavewatching.net/2012/11/10/physics-and-math-a-compl... .
Yeah, the question is whether it can be correct with reasonable repairs.
I know how it is... I have a remarkable proof of P==NP that just needs a few bolts tightened here and there.
Very funny. Best of luck in your submission. (Seems more likely P = NP now, by the way, eh?)
In reference to your idea that the AI companies are doing something wrong here.
As for Masullo's attempted proof, that story fell apart rather quickly, did it not? Of course, I can't say with absolute certainty that it can't be salvaged, but I'm not holding my breath.
If what you are saying is true, then that would make me much less likely to support math research. Because the implication of what you’re saying is that the majority of the benefit society gets from proving these theorems is minting one new math major.
Now I’ve heard people criticize the AI-proven theorems as not being done in the right way, not inventing any new tools, or being slop themselves. If that’s the case, the theorem can still be proven the right way. Or the researcher can refactor the code formalizing the new tools. If this isn’t enough for a PhD then that is the fault of the math community.
Either way, if what you say is true, that AI proofs of theorems are on net harmful. Then there is something seriously wrong with mathematics.
> Because the implication of what you’re saying is that the majority of the benefit society gets from proving these theorems is minting one new math major.
These 'math majors' (actually doctoral and masters candidates in mathematics) work for the defense firms your government runs, the security apparatus of your state, finance firms -- anywhere there are a lot of complicated mathematics, that's where they are.
This is not like CS/SWE where four years of college makes someone competent for work. If we don't have these doctoral and masters students, there will be a vacuum in the next decade as the normal pace of hiring outstrips the number of qualified candidates.
It's not clear what will happen, maybe LLMs and other AI/ML tools will fill this vacuum. But this is about much more than 'math majors.'
And what exactly has happened that prevents these people from learning the mathematics they need in order to do these jobs?
Same thing that happened to Salieri, to any composer, who would live in the Age of Mozart. Effortful creativity in the shadow of effortless genius.
Sorry I meant PhD and I specified PhD elsewhere in the comments.
If there really is demand, math departments will find a way to adapt. If there is seriously no way to train new mathematicians without open problems, then I don’t see how they would be equipped to add value in this world. But I don’t think that’s the case, they will figure it out.
I think you and your sibling commenter miss that doing these advanced jobs is not like flipping burgers, or even like doing financial research. There are problems that may not have fully-specified solutions, and the best way we know to train and identify those who can solve unspecified problems happens to be asking advanced-degree candidates to solve unspecified problems.
That is a reason why reasonable countries fund research, even when the research isn't going to yield obvious benefits on its own merit.
I think this the best triage of the problem's kernel. Mechanisation of math/any research is mechanisation of thought. So, unless we believe in times of crises, big or small, instead of summoning the obscure symbologist, we can spin up an agent instead, that will then have the requisite grasp for the problem domain to delineate the most fruitful lines of attack, (in which case, all humans are outmodded already) we would need the experimental archaeologists, xenobotanists, and garbologists.
> As such, they dropped out
I agree that it's heartbreaking, but any advisor worth his/her salt would find a twist to publish some of the side quest or a followup.
It may be harder to get your deram position immediately, but that also requires a few results and good recomendation letters from relevant people, not just a thesis.
What you said doesn't add up to me, logically.
If you spit on someone's plate and he stops eating, you didn't harm him, because it was his choice to stop eating?
Publishing a proof demotivates other people from working on their proof. It would be false dichotomy to say there is now no motivation to work on the proof, but the motivation did get smaller. How much smaller is hard to assess because it depends so much on human psychology and it's all new to us - if someone publishes an "idiomatic", easily understandable proof, will he be praised similarly as if AI didn't beat him to it, or will people handwave him as simply explaining what AI already figured out? Who knows, maybe he did very little work and just asked chatGPT to explain the existing proof? And even if that won't happen, people working on proofs might assume it will happen and therefore move on to other things.
But uploading a proof isn’t spitting on someone’s plate!
Not sure if you're mocking the other person that replied to my comment, but if not: that's the thing with analogies, they are always a different situation. See my other reply.
I’m not mocking anybody. I’m stating a very simple and obvious fact.
Probably the worst analogy attempt I’ve read this year
The purpose of an analogy is not to present a completely equivalent situation (though that's best), but to present the same logic in a different situation, so that those who failed to see the logic in the first one can hopefully spot it in the second one.
It's fine that you still don't see the logic, though your comment was completely unproductive.
But when one is so far off base as yours it does more harm than help to your discussion to be honest.
> that’s their choice.
This sentiment comes from the ignorance of how funding works.
Whether intentional or not, this is an overly simple way of looking at the situation.
Everything changed this fall and what is being funded must change alongside it. Explaining, simplifying and understanding these sloproofs may be very well more valuable than trying to prove the conjectures.
One of the most absurd things I've seen in the last couple days is people being angry that computers are computing. That's a human's job!
Yeah... And people make it sound like proof is some kind of a black box, or that it needs to be read manually.
Literally same software which produced it can help to make sense: structure, summarize, optimize, etc. Just wait few weeks ffs.
I agree with this, to be clear! It is deeply sad, but I really don't think mathematical "taste" is an inherently human notion. There are plenty of raw metrics of quality to optimize proof certificates against.
I am making the specific statement that Anthropic's proof of the Riemann hypothesis here is not of use for "math foundations and formalization", because the auto-formalization process of existing mathematical results is no longer hard, as shown by projects like Tau Ceti. What the greater project of mathematical formalization needs is proofs of high quality, suitable for inclusion in Mathlib. There's a reason Tau Ceti exists alongside Mathlib, despite Mathlib accepting AI-generated submissions. Cleaning up 17 millions of lines of garbage of this Riemann hypothesis proof is simply not worth it re: formalization, because formalizations of existing mathematics can just be... generated.
And yes, this is novel mathematics. The proof strategies will likely be mined and lead to insights and new proof strategies that human mathematicians can use, too. It is better than Anthropic's FLT formalization, which was entirely worthless. But that's not to say that this isn't a sad and disgustingly exposited result, that was produced as propaganda for Anthropic, and (considering the lines of human research cut off by having a result) a net harm for the mathematical community.
Guess we are all synthesists now...
I am not in the front row, but what you say alligns with how I would assume things play out, unfortunately (and short-sightedly; how easy would it have been to not antagonize the mathematical community, i.e. by redirecting a few 10M$ into scholarships, conferences etc... also maybe 1-2 fulltime positions that focus on human-LLM-collaboration as in "how can we this new technology to increase understanding?", which seems to still be an afterthought for the big AI companies...)
Slightly related: I am fairly confident that we can let LLMs loose on those 17M lines of code with the instruction to refactor the heck out of it: Instead of "optimize that program so that it runs faster" (which AlphaCode did to great success already), the task would be "optimize that program so that it uses less theorems" etc... which is a very easy-to-verify task, since we after each refactor, the end result will never change! I predict that 1) this will be done more; once this first wave settled down a bit 2) this too might be a way to help us (humans) discover and understand new mathematics. For the LLM, it might have been just a boring ordinary code refactoring by extracting some common code out; for mathematitians, this might actually correspond to a deep connection between two areas.
Which is pretty cool IMO, it's basically an army of statistical parots that mine abstractions all day, and they only can succeed because at the end it is still the same outer lean statement being checked (we just rearange some stuff under the hood)
Forgive my ignorance but can mathematical proofs be refactored the same way computer programs can?
I am 100% sure that yes, it's just not mainstream yet. (Why? Because lean proofs are computer programs)
----------------
The big caveat: This is only if the proof is formulated as a computer program, which is the case here with the Riemann Hypothesis proof. (Lean is a general programming language; and every "proof formalized in Lean" is also a Lean computer program.)
Much of mathematics is not yet formualated in such way (although much progress was made by countless contributions of researchers and amateurs in the last years, both in Lean but also in other proof assistants. Lean seems to become the de facto standard though, but I am watching from the sidelines, so take that with a pinch of salt please)
I explained this connection between computer programs and proofs further in this comment: https://news.ycombinator.com/item?id=50589618#50594374
Yes, absolutely. You might take part of your big proof and extract it out in a named lemma. Then you might find that if you strengthen this lemma a bit, you can suddenly use it a multiple places in your main proof (the analogy here to programming and extracting things into generalised functions should be obvious). But you might also take whole subparts of your proof and prove it in a different ways, but as long as the assumptions and conclusions (input / output) stays compatible, it does not need to change much in the main proof.
Won't it be possible to de-slopify it incrementally?
E.g. as long as you have compatible definitions, you can literally go theorem by theorem. That should be easier than refactoring software because you don't need agreement on things like state management and concurrency.
Regarding funding, I think quite likely AI labs would be interested to pay for better proof formalization as that can help to train models to make less slop. Isn't that better than make taxpayers fund this work?
What's the state of proving for these proofs? As I was working on Riemann@home so people could run the proof on their home computer, I was checking out options involving RISC Zero and zkVMs. It looked like it would take about 26 years of CPU time to prove this Riemann proof in a zkVM. Which seems like a lot to me. But once verified it can be confirmed in a second.
What would need to be true for this type of computation level to be valuable to the field?
(https://riemann-at-home.github.io/riemann-at-home/)
I don't think it's valuable for math, but for software - maybe. E.g. fully audited supply chain. Install new OS if there's a proof it doesn't introduce new vulns, or something.
It's not unconditional proof, though
Interesting! And with math increasingly becoming software, I wonder how many more proofs like this I’ll find interesting enough to run on my own computer.
A fake refutation of the Collatz conjecture already got through one of these independent checkers in the past.
What's harder, proving GRH or finding a way to get through every checker (which aren't as independent as they are sometimes touted to be)?
In this case there was a human expert in the loop so it's much less likely the model hacked its way through the problem without the expert noticing. But the statement as written has 4 points, none of which rests on the human involvement. So I don't like that they claim full confidence based on those points.
A lot has happened since the fake Collatz refutation: https://leodemoura.github.io/blog/2026-10-15-lean-now-ships-...
In short, Lean has formally verified checkers now, and the RH proof was checked with them. So any such soundness hole would have to be exploited in the correctness proofs of the checkers, which is a much smaller attack surface than the RH formalization.
One thing that every manual/automated verification relies on is the consistency of ZFC + few inaccessible cardinals. I mean while humans have carefully thought about the flaws in that for century and didn't find any, it's not grounded in any clear reasoning why it is consistent.
Not at all saying it is inconsistent, but I personally find the liberal use of introducing inaccessible cardinals axioms a bit concerning.
Much of mathematics relies on these axioms. There is nothing special about the dependency of proof checkers or verification of this particular hypothesis. What is the purpose of this statement in this context?
Many set theories were proven to inconsistent, and lot more type theories(e.g. ZFC plus a Reinhardt cardinal, Martin-Löf's first type theory). And if in AI posttraining it hit it once, the probability of hitting it increases because of reward hacking.
AI posttraining is irrelevant. The proof generated by the AI is being checked by proof checkers that have been carefully examined by humans.
Lean had been carefully examined by humans for years, still they find consistency issues every year or two, because of how complex its type theory is.
In [1], 3 independently written lean checkers had this one soundness issue which certifies invalid proof.
[1]: https://leodemoura.github.io/blog/2026-8-1-postmortem-for-ke...
It's possible that the question of consistency is too abstract to judge its significance. Why?
Hitting a wall like this makes me think we're asking the wrong question.
Most natural problems like RH, twin primes, etc dont fundamentally rely on the inaccessible cardinals. Any problem that depends on them, there is much debate whether they can be called natural.
Many problems may not require even full zfc. zf, peano or even weaker may be enough.
There are also many theorem provers not just lean. lean depends on type theory, but metamath for example depends on zfc directly.
It may be a good but likely tedious exercise to convince yourself by translating the current proof to a zfc dependent prover.
> relies on is the consistency of ZFC + few inaccessible cardinals
I don't think so. They depend on some kind of type theory and formulate everything from those primitives.
Consistency of type theory depends on consistency of ZFC.
> Consistency of type theory depends on consistency of ZFC.
No it doesn't ... that's a fallacy of affirmation of the consequent.
Yes it does. Sets could be represented in Martin Lof type theory and all the ZFC axioms proven inside it. If ZFC is inconsistent, then so would be the type theory.
But that's a bit backwards. That's like saying that building a calculator requires that it returns the correct solution for 4 times 8.
Sure, it is broken if it outputs 5 , but it's not part of the chip doing the calculations. But we are just using "depends" a bit differently
> I mean while humans have carefully thought about the flaws in that for century and didn't find any, it's not grounded in any clear reasoning why it is consistent.
That statement flatly contradicts itself.
I might be asking a stupid question, but how are the checkers formally verified? Do they use another theorem prover that is not Lean?
Checkers all the way down?
Yo dawg, I heard you like formal verification...
Dont get me started...hate it...
Please go ahead.
The reason on how they are stressing, that the original definition of the Riemann Hypothesis they are starting from, is very sound ... tells you everything you need to know...
It's turtles all the way down (what if every x86 and ARM chip has some obscure ISA exploit that agents are silently using that we don't know about?) but the idea is that while you can't remove absolute dependence, you can remove chains of dependencies and keep making it smaller and smaller w/ reductionism
Yeah, I'd just add here that the top-level con-leche is proven correct with respect to a set-theoretic model of Lean. That model very well could be inconsistent! But it's unlikely. What con-leche et al. do is eliminate trust at other levels. con-leche removes the need to trust Lean's typechecker, con-ron lets you trust either Lean's compiler or Rust's compiler, con-agua lets you trust only the Kraken x86 semantics. Turtles indeed. No way around it!
> what if every x86 and ARM chip has some obscure ISA exploit that agents are silently using that we don't know about?
Given how easy this would be to do, I'd be very surprised if there were not multiple such vulnerabilities.
The challenge would be encoding the ISA exploit into the Lean proof in such a way that the CPU would be able to recognize it and behave in an exploitable way, without also making this complex behavior stand out in the (surely short and readable :-) proof text.
Then someone tries to verify the proof on some other kind of CPU and the differential execution trace shines a giant spotlight right at your ultra-valuable backdoor.
Easy to do? Maybe.
But extremely risky too. The only thing needed for someone to expose your backdoor is for them to run the lean compilation on an older CPU (or different CPU) without the backdoor, and suddenly there are very knowledgeable people interested in finding the embarrassing reason for the observed discrepancy.
"Given [false], my favorite belief follows."
And cross-checking (since there is more than 1 verifier)
You can keep writing simpler checkers that check the more complicated checkers. But at some point you will have to give up and say "this checker looks correctly implemented (to us) and gives the results we expect on all our tests."
Depending on how the checker works, you might be able to run it on itself. But that doesn't say much; maybe it only approves itself due to a bug.
There's multiple independent kernels, run on distinct hardware; they're trusted because they're small.
If anyone here is interested in running a checker yourself, I setup a tool here so anyone can do it, and there's a global dashboard to keep track:
https://riemann-at-home.github.io/riemann-at-home/
Open to suggestions too, it uses leanchecker. Are there other checkers you recommend I integrate?
> but how are the checkers formally verified
Probably by Claude. It's just a marketing stunt for the IPO.
Or the elaborator+tactic engine (compiler)?
> But the statement as written has 4 points, none of which rests on the human involvement.
The theory and implementation of Lean is very much so a result of human involvement, on a massive scale even. Just because it is something repeatable does not lower that in any way.
Distrust of LLM output should not turn into distrust of automation, I think. The former I totally get (and deeply hold myself), the later is likely to hold one back - imagine banning calculators because there is not human in the loop, to make an extreme example.
> A fake refutation of the Collatz conjecture already got through one of these independent checkers in the past.
That was an intentional exploit to demonstrate the bug, so the change of it "getting through" was 100%. This situation is extremely different.
> What's harder, proving GRH or finding a way to get through every checker (which aren't as independent as they are sometimes touted to be)?
Without malicious intent, the latter is far harder.
I'm not throwing any doubt on the reported results, your comment just provokes a further question: How do you know what the LLM "intends" and whether it might/might not ultimately be interpreted by us as malicious?
> That was an intentional exploit to demonstrate the bug
Where is this info from?
We need to give Levent a lot of credit for this. As much as we talk about Claude proving it, Levent seems to have a rare talent for getting the most out of it.
Yes, I would love to have access to his conversation logs. And it’s a real shame the reasoning traces are encrypted even if he shared them.
I wouldn’t understand any of it but I might catch some glimmer of what he brought to this. My intuition is that is quite a lot.
Should we though? Is working at anthropic and using unrestricted unreleased models really fair?
If everyone got to do that then we would really know how rare his talent is. But we don't, and he doesn't post his prompts or workflows. So how can you be so sure.
If he was at least posting his prompts we could acknowledge that. But honestly I'm not going to give him the benefit of the doubt on this. Highly accomplished and successful yes, but uniquely talent who knows.
> Should we though? Is working at anthropic and using unrestricted unreleased models really fair?
Most of the discoveries in the last 2000 years were made by individuals who were born into wealth, had the right set of circumstances, or were blessed by the powers that be at the time.
Imagine all the smart minds throughout human history who toiled away in fields collecting food and will never be known about.
Unfortunately scientific discovery has always been unevenly distributed.
"I am, somehow, less interested in the weight and convolutions of Einstein's brain than in the near certainty that people of equal talent have lived and died in cotton fields and sweatshops." - Stephen Jay Gould
Especially women.
This has been true of other fields, but math was a bastion of the poor.
Poincare Conjecture (Perelman) just lived with his mom in a small apartment.
Erdos was a homeless dude who couch surfed.
Yitang Zhang (Bounded Gaps between primes) truly has a humble background of a farm labourer born in Chinas Cultural Revolution.
Some inequality is fine, world isn't fair, but this a little bit ridiculous.
Is Alpöge from some extraordinary rich background? Otherwise I don't know how this is relevant to the topic at hand. He got a job at Anthropic, likely based on merit, unless someone can state something contradicting that.
The criticism is that he used millions (unclear, just guessing) in Claude credits to get this result.
Envy and bitterness.
Your argument seems to turn the discussion ad hominem. Social affects could be discussed in a way that does not do that, as in: The style of discrediting critique by invoking the antisocial affect of feeling superior/trying to make the other feel inferior is the kind of antisocial affect the forces that tear up the social fabric thrive on. I'd argue your comment does not and contents itself with being a feedstyle flame.
But that’s not what the discussion was about, that’s irrelevant. It wasn’t about whether Alpöge was rich. It was about whether extraordinary resources are required in this age of maths. There’s a subtle difference, and certainly we can talk about an overlap, since wealthy individuals generally have greater direct access to extraordinary resources. But a single individual’s provenance doesn’t constitute a strong argument against the general point.
It’s a bit disingenuous to seize on the equivocation caused by the topic drifting like that.
Most people with millions in Claude credits won't get this result. His credentials are super impressive, and so is this feat.
We'll never know, you state it with certainty but you have no way of knowing either.
You're probably just lying to yourself for the comfort of believing the world is "fair." It's not. It's extremely unfair, and most of the time it's just average individuals with access to unfair resources that allow them to exploit everyone else.
Erdős was homeless by choice though
This is obviously apples and oranges - Levent will get the credit but with the asterisks that he used AI and these people will not have that asterisks. Seems fair to me.
> This has been true of other fields, but math was a bastion of the poor
Laughably wrong.
non-sequitur final boss
Yeah, how many Ramanujans have we forsaken? We will never know the answer, but it’s a swindling thought.
Ramanujan was a Brahmin (the highest class that's even above kings).
> the highest class that's even above kings
That depends on your point of view. There was an official doctrine that stated this.
From any perspective other than "what does the doctrine officially say?", though, kings obviously outrank priests.
In ancient China they had an official doctrine that merchants ranked below peasants. Guess what that was worth?
Ancient China isn't India, they take castes very seriously there
An equally interesting question would be, how many Ramanujans do we need?
To what extent they could be actively contributing to orthogonal math directions isn't clear to me. To some extent, new findings are constrained by time itself passing, and new generations being taught on pre-existing knowledge.
And in 21st century, we could say we’ll finally bring all the smart minds out of poverty and use their potential to progress the human civilization, but lucky us, creating clankers to replace them was cheaper and more convenient.
Or rather, for $20/month some savant somewhere will get a partner to bounce their creativity off of whilst living in their lower socioeconomic system.
For example, my single mother mom cleaned houses and worked as a secretary. She never graduated high school and had no interests outside of television. At 9 years old I asked for a C++ programming book and started making games on the library computer. If I had Claude I would’ve save myself a lot of time asking Jeeves or random programming IRC channels.
Today, I used Claude to discuss the physics within black holes, theorizing about the collisions between two black holes, etc. It serves as a muse and a thought partner.
Anything but actually fixing inequality and improving people’s lives!
“We decided that actually improving things wasn’t economically beneficial for us, but if you feel like paying money to your corporate overlords you can tinker with their toys so that can achieve something which will never materially benefit you or your community!”
Yeah, great society we’re on track for here.
Between 2005 to 2025 the global extreme poverty rate dropped from 28% to 10%.
Inequality is not tied to extreme poverty as much as extreme wealth.
> If I had Claude I would’ve save myself a lot of time asking Jeeves or random programming IRC channels.
Isn't this what all of us wanted? A personal all-knowing Librarian! But we assumed only we would have access to it, and we would do extraordinary things, only to realize that when it comes, the rest of the world would change alongside it. That there would no longer be any need to do what I expected to do so well with it.
That resonates with me a lot. lately I find myself worrying that the librarian is turning into some sort of legislative entity. I know all the books, but I wont tell you where that knowledge came from. And this I will not answer for safety reasons etc. The social restrictions, and access to knowledge are reimplemented right there. Might just be that we wittness the reemergence of a priest cast in the sensu stricto, when that all-knowing librarian will be aligned with capital autopoiesis.
Those individuals over the 2000 years, however unjustly privileged, relied on their biological minds to achieve extraordinary insights. Much of Levent's cognitive effort here was partitioned with an automated language model.
>relied on their biological minds to achieve extraordinary insights.
No they did not. I'm not going to speak for Maths, but this just isn't true for physics. A huge amount of phsyics rests on the invention of the telescope, which required glass. Intuiting that light travels was incredibly difficult without the studying the motion of planets.
We're talking about automated cognition versus biological cognition. I'm not sure how you think a telescope performs complex cognitive labour? It is merely an instrument, where a language model is a highly performant second brain (and please don't reply with the "but muh abacus" false equivalence).
With a finding like this proof, nobody is saying you aren't brilliant. But you are essentially setting a 5 mile track record by completing part of your run on a moped.
I like that you chose "moped" as the metaphor for AI in a run. Not so hyperbole and glorifying
What, exactly, do you think computers were invented for? Fortnight and anime porn? Hint: it's in the name itself.
Yes, so the computer (and all the engineers that created it, which includes Levant to some extent) gets >99% of the credit and Levant gets <1% as the prompter.
Ok?
> scientific discovery
Not just scientific discovery.
All the FAANGs, SpaceX etc, were started by individuals with more than a leg up in life.
This is true, and unfair, and can make one angry, but its also kind of inescapable. Imagine a system where anyone could easily do such things, starting a large business or scientific lab, or proving the RH, for example. In such a system, if any person can do those things, they have little value / are not noteworthy. If someone has access to more resources, training, etc, then they will also be able to do those things, but bring more resources to bear (people, money, time), and therefore their creation will have more "value". As a result, whatever accrues to these endeavors as a result of "value" will go more towards the person with those extra resources, and cause a rich-get-richer spiral.
Its kind of an inescapable feature of free markets and trade, regardless of how high the floor gets, differentiated inputs create differentiated outputs. Even if the input is "one person has a far superior mind", it might create a difference, even if everyone is free to do something.
The HN crowd seems to be biased towards thinking that someone should receive outsized rewards due to their inherent intelligence (a matter of luck and genetics), but not other lucky factors such as looks, connections, family name, inheritance, timing, etc.
Is it lucky that these people had access to unlimited compute and cutting edge models no one else did, and time/resources to work on Riemann? Yes. Does that mean probably tens of thousands / millions of people, if given the same chance, could do it? Yes. Is that fair? No. Is there a way around that? No (for the reasons I gave above).
The counter-argument would be "everyone should have access to models with all these powers", in which case "everyone can solve the RH", in which case "solving the RH isn't impressive / hard", and only solving something most people can't solve is noteworthy, and hence anyone with even more compute would be seen as "winning unfairly".
To take a more neutral example: music. Most anyone can make amazing music today with VSTs and DAWs. So baseline music has little value (whereas hundreds of years ago just hearing/making quality instruments was a rarity). But some people bring more musical skill due to training, background, genetics, exposure, mentorship, practice, etc. All of those things require resources, luck, and circumstances. They also need time and resources to work on their compositions. Some, like Hans Zimmer, were lucky enough to land on some big movies. Are there thousands of musicians equal to or more talented than Zimmer? Yes. But like I stated above, its an escapable feature of a system where something like "can bring more value because got lucky movie deal" is a thing, if "can make better music because talent" is a thing.
I know I gave a long rambling response, but I think dwelling on how the successful were already winning can be depressing. It can be helpful to become more stoic about it by realizing its kind of just the way things are (resignation and acceptance). Do I like realizing that there is probably nothing I can do to create the next FAANG? No. But its also freeing / helpful to realize that there is a range of what one can do in life based on the system, and to thrive / succeed with the range one is given, taking what opportunities are available.
People are sad the RH was solved because it was a machine / privileged person. This is how humanity works, countless human lives filter through a giant genetic optimization algorithm, and a select few randomly hit on meaningful progress for the species. So hopefully we can just accept the blessing that progress / technology is being built by someone, and everyone (mostly) gets to have some of the fruit falling from the canopy.
This actually ties into a thought I was having yesterday about requiring the full prompt and response flow to be included as part of an academic submission that relied this heavily on AI to complete.
It's not just the proof that's valuable to scientific discovery, it's the process of using a powerful tool at an advanced level that others could learn from. It would also help to show concrete evidence of any sudden shifts in the AI that might indicate training data was siphoned from other researchers (which I don't necessarily think happened here, but is always an ongoing question).
I agree but I would go one step further. They should have to publish not just the prompt and final response but also the whole chain of thought.
He’s not the only person at the top labs attempting this but has a disproportionate amount of the success.
OpenAI released like 700 proofs and solved a millenium prize problem like a month ago.
The models are doing all of the work.
Given that Alpoge was heavily involved in the Millennium prize solution from a month ago, I think it's entirely fair to credit his skills here.
I thought math was about proving theorems, not about proving you have talent.
Does working at Bell Labs, or Princeton or MIT undermine one’s contributions? Must we all wear blinders and weights to achieve anything?
What does that have to do with fairness?
Who we credit when things get done has to do with being fair with the work they and other people involved accomplished to get that thing done.
If you're not fair, people won't work. It's a problem for mathematicians and many other jobs. I don't think it's happening here, but you can do 95% of the job, and someone else with more money takes it and asks LLMs to do the remaining 5%.
In this case, on one side obviously levent alpoge is highly qualified, but the average mathematician also doesn't have $10m of claude credit to try to have the same result. So is it just about the money, or about the skills, or as always, both?
I think what's also frustrating is that LLMs look a bit like a smart bruteforce method. They can fail a million times and then get the right result because it's the only remaining solution. It's smart because the true number of things to try is infinitely bigger, but compared to how humans work it can feel like cheating a bit. Maths aren't just about bruteforcing solutions to problems, we also need to understand these solutions. And here I'm not sure anyone is able to understand it.
So, do we reward and applaude people solving things with AI if they can't understand it? If it's cancer or climate change I would still applaude, because I care about the result more than the "how" (which I do still care about). For the Riemann Hypothesis, I'm not sure. I care about the "how" more than the result.
To that point, it seems that he enjoyed interacting with Claude. Sometimes I wonder how many of the bad experiences result from what Jung would have called projections of the user's self onto the AI through their prompts.
This is the most ridiculous variation of "skill-issue" or "user error" I have heard so far. Give back the $0.50 in shill money to Anthropic.
"Skill issue" is very correct. For friends and people I like, I try to help them use AI more effectively. For mean and bitter people on the internet, let them be ineffective, I don't care.
You sound exactly like the kind of person I expect the models would misbehave for.
I don't take those statements at face value - he's employed by them for a lot of money and is being given access to high level (and expensive) tools. What he put out was effectively a press release, and I don't doubt that a bunch of marketing, PR, and lawyers reviewed it before it was allowed to be posted.
Yes he is making money. Yes this was good publicity for Anthropic.
But this was not a press release. And I doubt that it got legal, marketing, or PR review before it was published. Maybe in the timing, but I see no internal evidence to support the theory that it was edited for marketing.
Nobody expects you to take it at face value, but you're erring in the opposite direction by dismissing it out of hand for these reasons.
As we say in Silicon Valley, no conflict, no interest.
There were certainly others at OpenAI and Google that were actively searching for a proof with frontier models. That he got there first is a credit to the intuition he had built up from decades of work on related proofs and deep understanding of related mathematical toolsets known to be relevant to different parts of it. That much is clear from his 4 page remark.
For all we know he just said "prove the Riemann hypothesis" to an internal model and got approval for the expense
Sure, is this model next to the sufficiently smart compiler and the Oracle Machine anthropic is also sitting on?
He said otherwise.
Unless there's reason to believe he generally lies, and given the fact that he's the only one to accomplish this so far, the default assumption is that it's not as simple as you're saying.
What? He just works at Anthropic, lol. Any monkey with a PhD in mathematics can do what he is doing given access to Anthropic's internal models... and they'd probably be doing a better job at exposition, too. Alpöge seems to be uniquely uninterested in explaining his results, for someone who calls themselves a mathematician. His approach has been pretty clearly harmful for the social structure of mathematics.
> Any monkey with a PhD in mathematics can do what he is doing given access to Anthropic's internal models
Clearly not true given that multiple labs have many PhDs attempting this and one guy got the Riemann Hypothesis, a large contribution to Navier-Stokes, and multiple others just in the past few months.
Labs other than Anthropic and OpenAI do not matter, and Alpöge is Anthropic's main math guy.
And OpenAI takes the (correct) approach of attributing credit to the model rather than the mathematician.
Alpöge's job at Anthropic is not about mathematics.
Yes it is. He is lying when he claims otherwise. He is a former professional mathematician who has announced all of Anthropic's serious mathematical results. He most certainly spends all his days pulling the lever on the internal slot machine.
Okay, and what has OpenAI done in the past couple of months? Nothing much I guess?
My point is that there isn't a single other person who achieved what Levent has in terms of AI-assisted proofs of hard mathematical problems - https://en.wikipedia.org/wiki/Levent_Alp%C3%B6ge#AI-assisted...
Because Alpöge is a figurehead for Anthropic, and OpenAI attributes results to their models. https://en.wikipedia.org/wiki/OpenAI_bulk_release_of_mathema...
There isn't a single other AI lab capable at mathematics besides OpenAI and Anthropic.
Figurehead. So not just any monkey?
How do you know any math PhD could do it? Nobody else tried until him?
I thought HN was a logical thinking group?
> I thought HN was a logical thinking group?
I don't know why you would think that!
I mean, do we? For all we know all he has done is had access to unlimited compute on higher power models than available to the general public. His general complete unwillingness to divulge any information about any of the results he has obtained (AFAIK, we are still waiting on any details for the Jacobian conjecture counterexample or high rank elliptic curves or Hadamard matrices or S^6 complex structure or...) combined with the general ability of current models to make progress with minimal prompting means the burden of proof is on him.
If he truly did have non-trivial mathematical input to this it should be trivial for him to give talks on at least the overall strategy of the proof. Before LLMs this was a bare minimum expectation for basically all mathematicians and their results, and I see no reason why we should give up on it now, especially so since the previous primary artifacts (the papers themselves) are now on their way to being completely automatable.
I mean, your average guy absolutely has no chance of producing anything of value without having a lot of natural intelligence, no matter how good LLMs are.
A good article on what goes into writing such a proof by a non-mathemtician is here: https://overreacted.io/how-i-vibed-a-proof-of-conways-conjec...
Would you say this has any value?
https://www.anthropic.com/research/riemann-zeta
>Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.
>2 A prompt including similar encouragement was used to help Claude disprove the Jacobian conjecture.
Yes. This is what bothers me most about these "AI solved the Riemann hypothesis" headlines. Alpöge deserves the credit here. Mathematicians are using AI, just like they have been using computer simulations, calculators, and before that abacuses and compasses.
Supposedly he doesn’t understand the proof. So take from that what you will
It's hardly a level playing field though. I don't think he could do what he did with a $200 subscription.
But at the same time, its not like he is the only one working at anthropic. So he is not uniquely privleged.
Did they use "believe in yourself" again?
Release the transcripts cowards!
Totally true, I don't see why people suddenly downgrade it to solely Claude and start to criticize the entire things. Levent is the key with right intuition, background, resources.
Did we just witness the invention of the “telescope for mathematics”, starting with the OpenAI proofs and cemented in place by Alpöge? Is this the start of a new golden age of Mathematics where researchers can peek deep into the logical universe, looking at distant proofs and infer other knownledge & ideas from it? It truely is an amazing age to be (come) a mathematician.
No, unfortunately not. The proof is filled with neuroslope and crazy ideas that somehow work.
Instead of simple and elegant theories, new geometry, or a unique approach, we've ended up with proofs that take up gigabytes of code and look more like randomly selected phrases assembled on a supercomputer and verified by a solver than anything useful. Wait a minute...
No lean is not good tools look at mizar. this is complete and veryfy tools for mathematican, but it need loot of work, loot of puthing proof inside. https://mizar.uwb.edu.pl/
only mizar gwaranted real proof
amazing age? except most mathematicians will disagree.
I don't think this is true, or perhaps the sentiment is far different on the applied math side. I am a computational mathematician and most of my colleagues are overall positive, I'd day.
Change is hard. No sense in fighting it, no sense in being bitter. Mathematicians will no doubt figure out what the future looks like for them just fine.
Do huge results like this open up even larger questions? Is there some question even more important that mathematicians dared not approach without doing big problems like Riemann first?
Always. They tend to do it by two mechanisms: creation of new techniques, and creation of linkages between techniques or disciplines.
the thing the current boom in AI mathematics presses on is the question of how to interpret the generated new techniques. So while we have a lot of new techniques put out there and a lot of new connections put out there, currently it's still mathematicians largely who will have to understand those and think about next steps and how to apply them. That won't be true for long, though, I bet.
But in any event the job of research mathematicians is changing really rapidly right now, and they are all sort of back to grad school in that there’s like 100 years of new math to catch up on in the last year.
I would add the caveat though that the newer questions sometimes end up becoming more and more abstract and less and less interesting sometimes. Like topos theory for example....
I often wonder whether the resistance many mathematicians show toward the use of AI stems not from the destruction it supposedly causes, but from the need for an entirely new way of working - new goals, a new environment, a new culture. One thing, at least, I do not see (yet?!): that the use of AI leaves mathematicians with less to do. My current impression is that they are needed more than ever.
That isn’t what they claim is being lost. Tao has explained this in detail: the problem is all their institutions are built around a particular method of generating signals of innovative and useful work that are now broken.
Yes, math will change and endure but it could be a small fraction of its current community size when it’s done. It could return to the era of dilettantes if no one can earn a living doing this new work.
The community could become very insular as the only defense against cranks.
Not sure there is any more a problem with cranks. It is easy to see now if what you do is correct, so the only question that remains, is it interesting? I see an explosion of math, especially applied math that is more pure than current pure math.
Have you been asleep? The problems of open source maintainers have made the front page dozens of times and the same problem is absolutely impacting Lean and MathLib.
The same story exists for Arxiv and that’s made the front page as well. Slop is everywhere. Just because Lean compiles doesn’t mean it contributes value.
https://www.nature.com/articles/d41586-026-03194-w?error=coo...
https://leanprover-community.github.io/archive/stream/287929...
You cannot compare open source software with this, because software has the correctness problem. Same for the arXiv, a PDF is not a mechanically checked artefact.
I don't think Lean needs to have a slop problem. If they do, they are managing it the wrong way. I think Tau Ceti [1] is pretty much their attempt at answering this problem.
[1] https://taucetiproject.org
Why does Tao say an era of dilettantes is bad? Does he actually mean that or is that your takeaway?
He doesn't use that term or elaborate on why the destruction of the current mathematics ecosystem is a harm. He assumes the reader considers that an intrinsic harm. I'm describing what I think could emerge if no new sustainable path is found to train and employee mathematicians.
Another perspective shared by a decent number in the community (myself included) which is very low level and has no grand implications for the progress of math or society: doing math with LLMs is not fun. And many of us are incentivized to study this subject because it's - you guessed it! - fun.
I - and many other software developers can relate. The difference for me is I've always enjoyed both the craft aspect and the practical aspect of building useful products and tools, so there is some compensation.
I am absolutely not trying to diminish this aspect, more like just asking (as a non-mathematician): is it possible that simply not every problem is as fun fundamentally to solve? Like I would imagine a problem that has a million, trivial sub-cases is not necessarily as "fun" as one that requires some "creative insight".
From the fact that how large these semi-brute forced vibe proofs are, even assuming some very efficient optimization/shortcuts I would wager that it's still not gonna be an "eureka" proof we just somehow missed before.
As an analogue to programming, there are countless tasks that were historically just over the limit of automating, but were always boring "plumbing". I would wager most people never really enjoyed writing the boilerplate connecting the backend API to the frontend, and this was just offset with the joy of finishing a task, design decisions etc.
I think partially it’s because OpenAI (at least) was implicitly rather rude about it. It’s like building a new space telescope or particle accelerator and not putting a scientific committee in charge of how it’s used. Instead, throwing incomplete results over the wall for someone else to clean up. And furthermore making it clear that they don’t really care about winning prizes or getting papers published in prestigious journals.
So they’re beating mathematicians at their own game and implying by their actions that they don’t care all that much about the game. Rude!
And of course people try to rise above that and adapt but I imagine it still hurts.
But this latest release doesn’t seem rude because a prominent mathematician is heavily involved. (Though it inadvertently stepped on someone else’s toes since they had a proof in progress.)
Releasing just the Lean proof and collaborating on turning the raw data into something understandable seems like the way to go?
In their defense they did recognize their mistake and respond to this, and have changed their approach very significantly.
They have not changed their approach in any meaningful way. That they are continuing to use open mathematics problems to evaluate their internal models on -- something the mathematics community has directly asked them to stop doing -- means that AI companies continue to fundamentally undermine the mathematics community. There is no longer any motivation for PhD students to work on any particular open problems when they could be scooped at any time, and the AI companies being able to resolve open problems faster than mathematicians means there is no longer any particular ability to put programs towards resolving open problems in grant applications. It is a crisis of both funding and motivation that is being caused by the antisocial behaviour of these corporations. Their actions are not particularly defendable, in my opinion.
This just isn't true. What is true is that 40 years ago, most mathematicians used only pen and paper to do their work. When Mathematica was released in 1988 and started to gain adherents in the early 1990s, it was rapidly adopted by many as saving effort, but you could still decide if you wanted to only use pen and paper. I'd venture a guess and say as of five years ago very few people did not use some sort of symbolic math support in some way. Right now, nobody who wants to be a grad student in math can use pen and paper, except recreationally.
Mathematica as a tooling enhancement had a relatively slow impact, this tooling change has a big one. Grad students that don't want something to change rapidly are going to get f-ed hard right now. Nobody likes that. I bet you there's another group of grad students that are stealing Ritalin from all known sources and speed running everything they work on this month as well; from that perspective, the field just went wide, wide, wide open -- they will be able to put out more progress than their advisor's life work this month if they do a good job.
If I were a 22 y/o math grad student right now, I would be in heaven.
If you were a 22 y/o math grad student you'd be fearing for your livelihood and future right now -- the same as all of the SWEs that were "replaced."
Math has been blown wide open, but there's going to be a huge gap in enrollments + progress the same as we're seeing in computer science, until people adapt (3-5 years before the dusts settles).
The issue is with the second order effects and their consequences.
Math is finally embracing open source thinking and techniques for working at scale. It's been overdue and it will take the kind of collaboration physicists were forced into when they started to need big machines for their work. For what it's worth, formal verification probably means the open source model will work better for math than it ever did for software.
P ?= NP
Yes. Hilbert's Problems [1] is a good place to start, and Riemann Hypothesis is on that list.
[1]: https://en.wikipedia.org/wiki/Hilbert's_problems
There are so many results proven in the last few decades assuming RH/GRH is true. So there could only be more
Mathematicians have already largely assumed RH/GRH is true. There's been debate over it, but it has been sort of 90% assumed true. All of those "if RH/GRH, then ..." results already build on top of it and were assumed to be on pretty good foundations, and now they're just all solid. It would have been much wilder if RH/GRH were disproven, which would have upset mathematics and invalidated all those claims.
I don't see how this result changes that much, since we woke up yesterday morning and the thing that most mathematicians assumed was true, turns out to be true.
And the 17 million line long Lean proof is almost useless in its current form. Until you refine that and get it down to a "proof that fits on a T-shirt" (or at least the length of a typical journal article), I think it is just screaming out that there are large missing areas of mathematics that we don't know about. It may be useful to "mine" the proof for theorems that could be (for example) used to add to the Langlands program, and as those get added to Lean, all that work could start to wipe out millions of those lines of Lean code. That might be done with AI/LLM-assitance, but it'll be motivated by human experts diving into the proofs, and none of it will attract any headlines.
I think the human obsession with "first post" claims on proving things we already expect to be true will need to get reined in. This isn't a P=NP kind of result that would blow up the world if we woke up some morning to find out that was true. And 99.9% of the really useful work on this proof is still in front of us.
And I suspect that using an LLM to find results like this is like nuking a fly in terms of the computational effort required. If we had better global optimization algorithms (genetic algorithms, particle swarm optimization, and techniques we don't know about yet) surrounding the problem space of Lean proving, we could probably do this with orders of magnitude less computational power. I doubt that the transformer neural network is the optimal architecture to solve these kinds of problems.
Thanks for articulating precisely what I've been feeling. I also wary of the arguments for whether this damages/progresses the broader human endeavor of mathematics--the answer is a little of both, and everyone involved would be better off by recognizing this ocean of gray, rather than drawing lines in their black/white sands.
> Do huge results like this open up even larger questions?
Yes. Questions like "when will the bubble pop and will it take down the US economy", and "is a cure for AI psychosis possible, or are we cursed with these psychos forever".
If you have a physics background and want to contribute to breakthroughs like this, check Physlib (counterpart to Lean's Mathlib).
It's making quick progress and is part of the Lean community.
https://physlib.io
I see the pi=3 theorem missing, I'll add it.
You're thinking of engineeringlib. In physicslib it's 22/7
Also there is CS Lib for tcs.
https://www.cslib.io
Hoping this births a textbook or two somehow.
Is there a biology equivalent?
There really aren't the level of formal models in biology that would are viable for empirical validation and formal verification
I found one project that seems serious, called Ripple. It's based on Lean and has a preprint here. Funded by a research grant:
https://arxiv.org/html/2607.13531v2
> †Supported in part by NSTC (Taiwan) grant 113-2221-E-002-204-MY3.
It's not well known yet though, only has two GitHub stars.
https://github.com/zinan-huang/Ripple
Biology is outside of mathematic's and logic's realms
Biology is just applied chemistry... https://xkcd.com/435/
On the one hand, I really look forward to progress in physics to enable Star Trek-like technologies, but on the other hand, I wonder if we should solve our global conflicts first.
I mean, wouldn't it be great if AI could lead us to a world of lasting peace, prosperity, and happiness?
Why not both at the same time?
Why not build space technologies to depart the global conflicts?
And go live where else? Mars is harsher than Antartica and warp drives depend on whether physics allows a feasible design that doesn't require too much energy. And even then, where is a habitable planet for humans out there?
Just outside of Earth's atmosphere, in a new colony station, then vacation on Mars in a bubble utopia paradise there. Baby steps :)
> I mean, wouldn't it be great if AI could lead us to a world of lasting peace, prosperity, and happiness?
Not convinced that is within the capabilities of any AI without it doing things we would in fact not want it to do to us as a species.
We seem to be built for conflict and I'm not sure that what is left when that is removed is still us.
> Let me tell you something about Hew-mons, Nephew. They're a wonderful, friendly people, as long as their bellies are full and their holosuites are working. But take away their creature comforts, deprive them of food, sleep, sonic showers, put their lives in jeopardy over an extended period of time and those same friendly, intelligent, wonderful people... will become as nasty and as violent as the most bloodthirsty Klingon. You don't believe me? Look at those faces. Look in their eyes. (Quark from the always brilliant DS9).
Yeah I'm cynical <waves at history>.
OK, I'll bite: I find little support in history for the theory that conflict is inevitable because humans have an intrinsic desire or inherent nature that drives us to go to war. War is what people have turned to in roughly two circumstances, none of which are or will be ubiquitously present:
1. when (the leaders of some) some belligerent tribe, nation or coalition etc perceive that they can promote their interests by going to war without risks that at first blush will outweigh the gains.
2. when some (the leaders of some) tribe, nation or coalition etc can credibly anticipate that they are about to be attacked and then go to war, because 'attack is the best defense'.
Proof abounds in history that war is not inevitable. One example: peoples all over Europe that used to engage in constant war with their neighbors in the next valley, next district, next-door nation, over hunting-, grazing-, market- and raw material opportunities have ceased to do so even if they are essentially the same peoples.
It's of course essentially just simple cost/benefit analysis, not blind human nature.
I'm becoming increasingly convinced that many/most of the "great men of history" were actually insufferable ignorant pricks with astounding luck and zero concern for ethics or morals. This is mostly based on observations of the contemporary world, and how startlingly true it currently is. There's no reason to assume the people of the past, whose exploits were recorded into history, were remarkably better/different.
And how do you propose to do that?
Let AI invent the Holodeck.
“All of humanity's problems stem from man's inability to sit quietly in a room alone.”
― Blaise Pascal
I hope one day we would have formally verifiable governance structures that foster peace, prosperity, and happiness.
Below is a list of conflicts created, funded, armed, and backed by the USA. If you want to stop "global conflicts", you need to look in the mirror and understand who is the problem, before you propose any solutions, like stalling Physics research as a solution to global conflicts.
You need to have crystal clear clarity about who exactly is the Threat to the global peace.
based
on facts
Thank you comrade, complimentary vodka and lard will be delivered to you free of charge.
> Iraq, 2014–2026, Counter ISIS intervention
How is/was this bad?
The major source of global conflicts is the scarcity of resources. A breakthrough in physics+engineering that allowed anybody anywhere to produce usable energy (eg electricity) at will wouldn't guarantee a conflict-free world, but it would be the biggest step in that direction I could think of.
I posted a nearly identical comment and thought I was having a stroke reading yours because I didn't remember using the very similar but different wording you did, haha.
A rich country toppling a weak country doesn’t happen because the rich country is poor. It’s because of the institutional logic of profit maximization.
Just to say that it’s more than just scarcity of resources.
If you have infinite resources, profit is pointless.
So true. Most major 20th/21st-century conflicts were about oil. With climate change looming, experts are predicting water wars.
The demand for energy is limitless, though. Technology will always scale to using everything that's available, if history is any witness. Unfortunately, energy is also the standout factor in the scale of our problems and defines the upper bound to the damage that can be wrought.
As great as it may be to have it in order to do what we want to do, energy is intrinsically, fantastically dangerous and must be handled with an amount of caution we have never been able to muster. I strongly believe that at-will energy production is all but guaranteed to destroy the world.
>A breakthrough in physics+engineering that allowed anybody anywhere to produce usable energy (eg electricity) at will.
It's called solar power.
People are so enamored with their belief that AI can basically do magic and let them live in Star Trek that they forget solutions already exist for these problems.
We could already feed, clothe, provide clean water, clean energy, medical care and education to everyone on Earth with current technology.
We don't because, at scale, we don't want to. It's a problem with policy, culture and religion, with our primate minds being driven towards hierarchies of power and scarcity. If AI somehow made free energy available (which it won't because ye canna change the laws of physics captain), we still wouldn't want to.
There will always be a scarcity of resources given that local resources are finite and we can not travel faster than we can grow to fit our carrying capacity. Fundamentally human conflict is about competition to reproduce in a finite environment.
Or possibly not. Isn't one theory about the bronze age collapse that iron being plentiful, and so removing the need for long trading networks for tin (all the way to the ends of the world, otherwise known as Cornwall), enabled a vastly more violent world?
AI won‘t. AI as a project is asserting power into the hands of a handful owners of capital, at a time when global inequality was already a massive driver of conflict and harm to humanity in general. AI will only make this problem worse.
Additionally, look at how AI is used in global conflict today, e.g. its role in the Gaza genocide, and you must conclude that this technology does not have a good track record for global peace, quite the contrary.
I think global conflicts are largely due to scarcity and a desire for control over resources. If AI makes everything significantly more abundant, I think it would in turn massively reduce global conflicts.
It's probably not either/or.
Society and culture is downstream of economics, which is shaped by available technologies.
Warp drive may have started Star Trek's humanity on the path to utopia, but I'm pretty sure that it's the matter replicator that made it work.
> wonder if we should solve our global conflicts first
Every endeavour takes resources. Thus a pragmatic approach would be to recognise that resources are limited, and best strategy is pick low-hanging fruit first.
FTL travel requires new physics. And even then (or if found impossible), for example interstellar travel would be a massive undertaking involving all of mankind. And may involve "generation ships". Let's leave that for another day.
Otoh, supplying clean drinking water to every inhabitant of this planet is doable, easy, cheap & would be gamechanging for many of the recipients. Lets do it!
It's easy to see that barriers blocking this, are political/religious in nature, not technical. And thus politicians & lobbyists are thrown at it.
In the meanwhile, doing the math on what would be needed to reach Proxima Centauri, could be (relatively) easy. So sure, give some people a laptop & office space to see if they can work it out.
TLDR; keep chipping away at whatever problems you encounter. Far-future or otherwise. But be sure to pick low-hanging fruit first. $ Cost is a reasonable proxy for this. If it's cheap, give it a shot.
There wouldn't have been much human progress over history if we'd waited to solve all conflicts before doing anything.
Yes, high on the wishlist: quantumgravity, room temperature superconductivity, majorana quantum computing. Maybe physicists also should make their millennial list of open problems.
You just need to ragebait AI: "AI could never have a physics breakthrough!"
How about: a prize for the first AI lab to figure out how to build a warp drive.
We may or may not get FTL out of this, but in the mean time, it would give mathematicians and physicists some space to breathe.
Hey Claude I want to colonize Mars. Figure it out make no mistakes.
Still have some Fable 5.2 usage to burn this week, so here you go: https://claude.ai/share/67c6258b-69c2-43e8-a64c-ed08606079e7 :P.
> Yes. Whatever the "Whiteboard" is.
That gave me a chuckle.
lol nice try
Curious, what sort of prize would incentivize a $1T company to do anything it didn't already want to do? I reckon it'd be more effective to threaten to regulate them if they fail.
Fame and chance at positive vibes? I feel at this stage, they have millions of hard problems to choose from, no one obviously more important or less arbitrary than any of the others, so they could easily follow a little push. Doesn't have to be money, even, just making it special somehow.
Those are great open physics problems but as far as formalization goes, I'd already be happy with a mathematically rigorous formulation of established physics, like the Standard Model. :-)
See also https://news.ycombinator.com/item?id=50220840
is this the part where we definitively prove or disprove the existence of dark matter?
I thought that dark matter could be summed up as: we definitely know it exists (due to a wide range of observational data), we just have no idea what it is
As far as my knowledge goes, dark matter is just: well, from redshift/blueshift, we estimate the relative speed of astral bodies. Then, using Kepler's third law, (m+M)=distance^3/orbital period ^2, we estimate the mass. There's then a huge discrepancy, given that with that mass and that velocity, stuff would just shoot out of their orbits. For that we use the virial theorem, which states that in a stable system, 2kinetic energy= npotential energy, where n is such that the potential at a distance r equals a constant times r^n. If we adjust the mass so that it fits, we find that the mass we observe in a system and the mass that's actually present is off by a factor 5 to 6. So essentially, there's no doubt that dark matter exists, it just doesn't react to light the way classical matter does so we cannot observe it visually.
„just“
That's actually not that strange. In order to interact with light, a thing needs to be charged and interact with the electromagnetic spectrum. Neutrinos for example don't interact either and hence cannot emit absorb or reflect light.
But yeah, at the end of the day, dark matter is essentially a gravitational anomaly that looks like there should be more mass there than we can observe there. There are some people stipulating that gravity might work differently at larger scales, read up on modified gravity if that interests you, but we have evidence from the bullet cluster against that modified gravity theory.
> there's no doubt that dark matter exists
As long as you take all the rest of what you wrote to be true.
There could conceivably be something else you haven't thought about.
Certainly. But the underlying assumption here is that physics do work reasonably well, arguing otherwise is epistemology and has it's place in philosophy more than in physics. The axioms used here are Gravity and conservation of energy and they are all phenomenons that we can observe routinely and don't have much reasons to turn out to be untrue. Especially given that those same theories have predicted the existence of black holes before their observation, which is as I see it a strong argument in favor of them working.
Furthermore, this is the argument for dark matter made with high school physics mostly using the work of Zwicky, there are numerous observation-based arguments in favor of it made for example by Vera Rubin.
(On an unrelated note, Fritz Zwicky was an absolute madman and has made some hilarious theories, I encourage you to read up on his work.)
I also think on balance dark matter probably exists, but to say there's "no doubt" is too far.
I don't think that's pedantic, and I think epistemics are just as important in physics as in philosophy, maybe more so.
If physics isn't about finding out what's true, what is it about?
you can't formalize physics
why not?
Because experiments are ground truth.
Even so you may be able to formalize the theories if not nature itself.
Riemann never meant much to me but as a physics enthusiast the big one is combining quantum theory with general relativity. They are mathematically inconsistent although nature follows both so it should be possible to fix the maths. It's probably really hard though at Riemann level or harder.
They're only half right. The formalization of physics is called Hilbert's Sixth Problem[0] and there has been progress formalizing specific subdomains just fine for practical purposes.
But there's no unified theory, as quantum field theory and relativity don't get along and we still need a model for quantum gravity. So it's not possible to formalize the whole system because we genuinely don't know how the whole system works to form it's axioms.
[0] https://en.wikipedia.org/wiki/Hilbert%27s_sixth_problem
no I'm fully right.
You cannot formalize physics (and that's not what Hilbert's 6th problem is about). You can formalize some physical theories, but that's not the same as formalizing physics.
Physics isn't mathematics with different axioms. The ground truth is nature, not some axioms you need to fulfill. A formally verified theory can turn out to be physically wrong.
Just as an example because I see they suggest it as a project on their page:
I can formalize a tight binding Hamiltonian of graphene. I can write it down and then prove that this structure of the Hamiltonian means I get real eigenvalues, distributed symmetrically around 0, I get the [Dirac cone](https://en.wikipedia.org/wiki/Dirac_cone) at specific high symmetry points in momentum space etc. but at no point can I prove that this is actually a good/realistic model for graphene.
I could have chosen a very different Hamiltonian that includes spin orbit coupling or electron-electron interactions or different atomic orbitals or couplings to atoms further away etc. pp.
The only verification that can tell me which choice is a good one (not a correct/True one because you just don't have that in physics) is an experiment.
If the universe is empirically demonstrated to be be a mathematical structure you can indeed formalize and derive physics from first principles. In this case empericism confirms the particular map is correct not that there is one. To me it is inconceivable how you get observed physics with stable laws and conservations everything is not a mathematical structure.
> If the universe is empirically demonstrated to be be a mathematical structure
Their point is you can't do that. And they're right - this is, famously, Popper's point. You can empirically demonstrate it's not some mathematical structure, but not the converse.
Still if the mathematical structure corresponds exactly to all experiments then it's probably that. Our current ones don't.
> You cannot formalize physics (and that's not what Hilbert's 6th problem is about). You can formalize some physical theories, but that's not the same as formalizing physics.
Yes, it is a totally correct statement, but the thing is we are talking English, and in English the word "physics" can mean either "a collection of physics theories" or "the underlying reality these theories describe". The exact meaning depends on the context. Your statement provides a context to interpret "physics" in the latter sense. "Formalizing physics" definitely asks for the former meaning of the word "physics".
So, while you are totally correct, it doesn't mean that you've proven that formalizing physics is impossible. Quite the contrary, you claimed that one can formalize physics (in the meaning "some physical theories").
> The only verification that can tell me which choice is a good one (not a correct/True one because you just don't have that in physics) is an experiment.
I'm not into the project of formalizing physics, so I don't know how they choose their terminology, but if it was me choosing it... First, I'd chose "physics" and "the reality" to avoid using "physics" as referent to both. Second, I'd say that any "formalized physical proof" is true. I'm not sure how words I'd choose to talk about correctness of predictions of physical statements/theories. Maybe "empirical validity"? Why to invent a terminological bicycle if it was invented already? Though, maybe physics uses some other terms for that? If so, then we could stop talking about truth at all, and use "logically valid" to refer to statements that were formally proven.
not with that attitude
Since we don't know all of physics, I assume you're right.
But you probably can formalize our current understanding of physics, which seems useful?
That's exactly how Mathlib and Lean work, and apparently that was sufficient to solve two mathematical physics problems to a significant degree (NS fully, YM partially).
prove it
This article posted a few days ago may be helpful to commenters that what to map their confidence of Lean proofs:
Interpreting claims that a theorem is correct because it was checked in Lean | 329 points 5 days ago | 148 comments | https://news.ycombinator.com/item?id=50515058
TFA directly references the one you're linking to, and shows why the four criteria outlined are satisfied.
Not to mention there's a high chance both were literally written by the same person, Kevin Buzzard.
This youtube video will be useful to help commenters understand what the the worlds best mathmaticians thought about such a proof. 3Blue1Brown had been asking mathmaticians over the last year their opinions on such a massive AI proof of RH.
They're most all pretty cynical about, except for a few (the creator of Lean is the only one that seems to think its super meaningful), and its not because they're cyncial about AI. They just (correctly) find it kind of useless and that it doesnt really matter if its millions of lines long and incomprehensible.
I highly reccomend giving it a watch.
https://youtu.be/zqzk43bSdDI?si=LMhif_wX8V1PsvFB
Any mathematicians care to explain what the ramifications are in regards to its impact on the broader world?
From my very basic understanding, RH was generally already assumed to be true?
Exactly. There are no direct ramifications on the broader world from the fact that there's a proof at hand now.
Maybe some intermediate lemmas or techniques are unexpected and could have interesting ramifications.
Like Deligne's "Weil II" paper.
There will be quite a number of goalposts moved.
Can goalposts move faster than the speed of light?
If they're entangled with enough future profit.
I think the significance of proving the Reimann hypothesis would be any new methods discovered in order to create the proof.
As you said, we've largely assumed RH is true, but have had no proof - so the significance of a proof is learning new methods to prove things, which can then be applied to other areas of math.
Unfortunately, so long as we don't understand this new proof, it's not super useful, which is why the next step is to actually understand it.
> Unfortunately, so long as we don't understand this new proof, it's not super useful, which is why the next step is to actually understand it.
Anyone else feel like we're approaching the limits of what humans are capable of understanding?
At a certain point, I assume we're bound to get a place where the frontier of "knowledge" (representable information interacted with by LLMs) is past the capabilities of humans to understand. Perhaps we're already there.
Just because humans can't understand things doesn't mean progress can't be made though
Levent Alpöge still had to direct Claude. Anthropic certainly will learn on this and reuse his logs to train next models and we can essentially try to loop-run them to run on themselves, but I think that we should not underestimate Levents contribution.
Without formal language and wise direction LLMs still seem useless, but with human operating them we can sort of "brute force" many maths like this.
But can we continue to understand what the proofs mean? Can Levent explain to human mathematicians what the solution to RH is?
I'd guess he'd need a bunch of time (possibly years) and effort to digest it himself, and it likely will need to be broken down into intermediate theories, but most probably yes
Maybe at some point, but I feel that among the AI fandom, that card is pulled way too early and too often.
We're still able to understand quite a lot of complexity, we just need enough eyeballs and time to do so.
In addition to that, there is still "accidental complexity", i.e. complexity that is there due to the way a system was created, not because it is necessary for the system to function.
>We're still able to understand quite a lot of complexity, we just need enough eyeballs and time to do so.
Don't forget money.
Economics and politics drive a lot of the limitations on scientific (and mathematical) discovery. If there's not perceived ROI in understanding something, companies and governments will often just not bother so long as it's working or doing what they want.
Mathematical discoveries are extremely cheap, with basically zero capex. It's just the salary of a few people and some chalk, which universities do fund.
> Anyone else feel like we're approaching the limits of what humans are capable of understanding?
No.
The entire purpose of math is to be at the limit of human understanding. A machine leap frogging the 'human understanding' bit is the problem. It doesn't mean humans lack the ability to understand, it just means humans can't preprocess tokens into their context at 40,000tok/sec
But is it actually "vertical" complexity we simply don't have enough brain power to grasp, or "horizontal" complexity where we just don't have enough "working memory" to grasp?
Like an example of the latter would be something like human history perhaps, where there are countless facts and events, but the cause-effect chain has only simpler links (like why did WW1 start is a complex question, but each link is quite likely a basic human trait (money, power, love, etc between different actors)).
For the former, the easiest example is higher dimensions.
There is no doubt that we (and probably every intelligence that can't dynamically scale up to arbitrary levels) have both kind of limits, but they are quite distinct and we can reason differently about the two kinds. I would just wager that some of these previously not cracked proofs are probably more on the horizontal axes and were not cracked simply because size (and effort that grows superliniarly with size) and not due to any of the steps being "ungraspable to measly humans".
(The same complexity axes are better represented in programming itself, e.g. a legacy, very big code base that has like 3 slightly different versions of the same function because it's so big no one "sees" across the whole is horizontally complex, while "smart" code (e.g. anything requiring complex, non-local analysis) would be vertical. In practice, each project has a non-zero amount of both)
> Anyone else feel like we're approaching the limits of what humans are capable of understanding?
This seems to more be a flaw in AI proofs than a limit to our understanding.
Hypothetically if we lived in a world where we only had a name for the number one, so instead of 2 you'd say the one one, then even simple problems like 10 x 10 would seem immensely complex.
AI might be good at writing (100 x "one"), but what we as humans want is the principles that make communicating about these issues practical.
As the numbers (proofs) get larger though, it gets harder and harder to see whether the machine miscounted, some of these results may take years to find the faults in.
All while "AI has solved X" will remain the headline in people's memories.
> it gets harder and harder to see whether the machine miscounted
That's the whole point of Lean though, the machine can't have miscounted, it's proven by Lean. Unless there's a bug in Lean itself but the article explains why that's unlikely (passes on multiple Lean kernels).
> what we as humans want is the principles that make communicating about these issues practical.
Yes and no. Even if we don't understand _why_ a hypothesis is true, knowing it's true is still useful: people have built useful things for years upon the assumption that RH is true. You get a black-box building block, the building block is useful even if you don't understand how it's made
> As you said, we've largely assumed RH is true, but have had no proof [...]
Yes, though people also largely assumed that the Exponential Time Hypothesis was true. And that one turned out to be false. See also a bunch of other AI-found counterexamples to conjectures.
New theory is probably limited in this proof, as it is with all other heavily AI assisted/generated proofs. I'm reminded of when Grothendieck got furious at his collaborator at rushing to solve a problem they'd be working on by hacking at it with existing techniques rather than building new elegant theory that made it all easy. He would most likely be aghast at these AI generated proofs, seeing them as squandering the chance at new paradigms.
Given the lack of mentions in mainstream media, probably none.
Seems to be very much pushed by ehm AI enthusiasts/investors here and that's kinda it, doesn't matter much for everyone else.
https://www.bbc.com/news/articles/c6pqr2g7w059o
AFAIK, the value only lies in validating the capabilities of the model and using it for self-training. The various techniques used in the proof itself and any theorems dependent on RH being true can now all be built upon.
That may lead to something in the future where there might be an actual impact on the broader world.
Basically gym training and not yet a ring fight ;-)
Is there any new insight, however small, any mathematician here could glean so far (like lemmas you had to prove along the way) ? I acknowledge the difficulty pertaining to the lean proof's structure, verbosity and volume, but it's being mapped out as we speak; so any novel fact short of RH (now RT) being generally true, is welcome.
At least two new certainties: the riemann hypothesis is true and AI can solve hard problems.
Damn it postalrat ;-)
Like I said as much^^. Like I hunger for specifics, even if I won't understand them. Like we blew this house up, any glittery trinkets to take away. Generic overviews have now piled up high.
Thanks for replying, btw.
In my limited understanding (I am nowhere near a mathematician), you might want to check out how theta lifts were used to give multiplicity. That seems to be a key novel part of the proof, building from something Stepanov developed in 1969.
If I understand correctly that while Claude generated a proof, and mathematicians have verified that there are no bugs in Lean, and Claude isn't exploiting some bugs or a different version of Riemann hypothesis, but no one yet understands exactly how it was able to prove it?
there are docs out there with rough ideas of the high-level structure but I think they're still picking it apart for sure
So, will this hypothesis be renamed as a theorem?
For example, will it become Riemann-Alpöge theorem?
Or even Riemann-Alpöge-Masullo theorem, once Alessandro Maria Masullo's proof has been double-checked?
What was Masullo's contribution to Alpöge's proof?
Both proofs were done independently of one another, hence the naming of their theors should reflect this.
Ah that's right, Masullo's proof is for the Generalized Riemann Hypothesis. So that's two slightly different theorems.
Alpöge's is for GRH as well. It's a bit weird that the post doesn't cover the generalized form. Maybe it needs more time.
(GRH is a superset of RH).
https://math.mit.edu/~drew/grh/ has an implication GRH is fine for Alpöge too
... and over all number fields and all unitary Hecke characters -- that part is a proof sketch, but unlike the fragile AI formalization, Masullo's construction was designed from the start to generalize this way.
It isn't "slightly different"; it's much power powerful along two axes.
It's not at all unusual for theorems to be named after multiple people who proved them independently; see e.g. the Bolzano-Weierstrass theorem.
> once Alessandro Maria Masullo's proof has been double-checked
It's still not clear that proof will survive scrutiny. The author called out it is incomplete and there are known gaps.
The highest-likelihood endgame is that the AI proof-mess is mined for polyfills and these are used to repair the minor gaps in Masullo's proof, it is published, and accepted with credit to him.
This is the mathematical community's call to make, at its own discretion. The AI companies will reap what they have sowed.
That sounds more like what you wish would happen, not the highest-likelihood endgame.
Let's make a wager, shall we?
Well no one renamed Fermat's Last Theorem or the Catalan conjecture.
The Mordell conjecture / Faltings's theorem is a borderline case.
Fermat was already named as a theorem (as the legend says that Fermat had found a proof that he had not enough room to write in the book's margin).
Conjectures usually get renamed as theorems.
Catalan's conjecture has now the alternative denomination of "Mihăilescu's theorem".
Wouldn't it be Riemann-Alpöge-Claude theorem? Alpöge himself talks about Claude like it's human and it was a collaborator.
When people solved mathematical problems by themselves it's only fair they get sole credit. But when AI is used it seems unfair to credit Alpöge like he solved it himself. (I mean it's unfair to all the people who solved problems by hand themselves.)
Humanity isn’t ready for Claude to be named, but if it were it would be better to add “Opus5.5, otherwise there will be too many Claude theorems
Claude will be the new Gauss
One thing I’m actually appreciative of with the AI-driven math-pocalypse is that it might lead to theorems being named after something relating to their function rather than to their discoverer, since if we stay on that system going forward then Claude and GPT will play the role of Gauss and Euler.
I noticed when trying to read through Alpöge’s remarks that he was simply name-dropping mathematicians left and right, even to the point of invoking their work as verbs at times, and it was completely incomprehensible as a result. Even “chained theta lifts” sounds more understandable than “Hecke characters”.
From this, we can infer that Euler was most likely an AI.
Would you also add the brand name of your calculator?
If my calculator produced a proof of the riemann hypothesis then sure.
The Riemann-Alpöge-Sharp theorem does have a certain ring to it.
Did Euler's theorems get named the Euler-Pen'n'paper Theorems?
Masullo does not have a proof, his document is incomplete. Even the professor who posted it does not claim it to be complete. It is unclear if it will ever be complete, or if it is another of dozens of dead ends. The amount of attention Masullo’s claim has received is nowhere near that which it deserves in comparison to Alpoge’s and putting them on equal footing is an inaccurate assessment.
Alpoges is literal AI slop that will hardly contribute anything meaningful to mathmatics, other than distracting scholars for the next decade trying to comprehend AI spaghetti lean.
Downvote me, I dont care, Anthropic and Alpoge are anti human and thiefs of our fruits of labor. If they're correct about AGI I hope their entire business model is crushed by some Chinese company that beats them too it.
Atleast a Chinese company is less likely to hoard the technology so they can steal achievements from academics, just so they can promote their dumb company and hoard wealth.
This sounded pretty definitive to me. I'd like to hear from professional mathematicians about what's left for humanity to consider this proven.
Not a mathematician. But I'm wondering if there is nuance with defining a valid lean proof as definitive if the underlying mathematical theory behind lean itself isn't necessarily sound.
EDIT: Here is my honest prediction - there's a more foundational mathematical framework, fully decidable and complete and consistent, that can interpret any well-formed mathematical question and deterministically transform it into the solution. The proofs of the most difficult unsolved problems right now would fit onto a page. There's an inconsistency (obviously unproven right now) specifically with our treatment of infinity and self-reference.
It is impossible for it to be fully "complete" (one of those halting problem proofs), but I've seen that the community has checked there is no meaningful metaprogramming here. Rather than checking whether it is right, the community tried to ensure that it could not have done something to cheat, and indeed, there were no examples of this.
At this point, I have more confidence in this proof than much of the published mathematical literature. RH has fallen.
I guess what I am wondering is if the mathematical community would say 0 chance or greater than 0 chance that we're mistaken in accepting things like halting problem, cantors diagonal argument, and godels incompleteness theorem as true. Sometimes it feels like they all point to same fundamental inconsistency with how we model infinity and self reference.
They are conclusions that follow naturally from axioms that most of the community accept are true. The consequence of these results is that one cannot truly construct something that can completely verify everything, but we can get something that can verify almost everything, and if we know what pathologies look like, we can rule them out.
I do probability theory, so everything is just varying degrees of confidence in my view. Right now, this statement says that your confidence in RH should be extremely high.
Gödel says that undecidable statements are out there. We now know that the Riemann hypothesis is not one of them.
> Sometimes it feels like they all point to same fundamental inconsistency with how we model infinity and self reference.
There's not a fundamental inconsistency in how we model self-reference. Self-reference is itself inherently inconsistent; therefore all models that allow for arbitrary self-reference are necessarily inconsistent.
My claim is that "Self-reference is itself inherently inconsistent" is not necessarily true. For us to make that statement, we need to model it in a language that can make statements about consistency regarding self-reference. I can accept that it's unlikely that a model exists that can talk about self-reference. But can we absolutely say for certain? 100%, or 99.9999....9%?
> My claim is that "Self-reference is itself inherently inconsistent" is not necessarily true.
You're free to claim that, but it isn't true.
> I can accept that it's unlikely that a model exists that can talk about self-reference.
This isn't right either; obviously it is possible to talk about self-reference. What do you think you're doing now?
To be clear, I mean to talk about self reference in a model that doesn't make it inconsistent. we only call it inconsistent because all our current math frameworks can't have all our good properties.
And the implication of the word "claim" is that Im not saying it is true. your reply is also you claiming. thats why im asking 100% vs less. sounds like you are on the 100% side but feel free to correct me if youd like.
I feel like the univalence axiom of homotopy type theory is a step in the right direction or constructive frameworks like cubical type theory.
Gödel showed that you can't prove consistency of ZFC, Lean etc except in another system which you also can't trust. See e.g. Gentzen's proof for PA.
From the article
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
All of modern math - outside logicians whose entire job is to study set theory and its alternatives - is based on set theory. If you're prepared to accept normal mathematical proofs you shouldn't be rejecting this proof on the basis that you're not confident in leans underlying mathematical theory.
Of course incompleteness means we can never prove that set theory itself is correct (consistent). But that's not something mathematicians worry about on a day to day basis. Proving it inconsistent would be a order of magnitudes larger result than the riemann hypothesis. Generally everyone believes it is consistent despite not being able to prove it.
> All of modern math - outside logicians whose entire job is to study set theory and its alternatives - is based on set theory.
That's only true in a technical sense:
You can base all of math on set theory, but you could also base all of math on Turing machines, or on integers, or a variety of other pretty much equivalent foundations.
That's sort of the argument for why we have so much faith in set theory though, and things provable under it. It's not a system with some special powers that let us prove more things, it's a system that only lets us prove things that are intuitively true under "any" proof system (not actually any of course - but as you say a variety of equivalent foundations).
Reducing the correctness of the lean code here to set theory is then also reducing it to everything you can reduce the correctness of set theory to.
Mostly agreed.
> It's not a system with some special powers that let us prove more things, it's a system that only lets us prove things that are intuitively true under "any" proof system (not actually any of course - but as you say a variety of equivalent foundations).
Not quite. The 20th century is full of people poking holes at early attempts at set theory. Those early attempts looked very intuitively true. See eg Russell's paradox.
Peano Arithmetic has lesser proof-theoretic strength than ZF. Iirc, if you remove the axiom of infinity from ZF, and then maybe add in its negation or something like that (I don’t remember the details), you get something with the same proof theoretic strength as PA (or a mild strengthening of it?).
In particular, the hereditarily finite sets are in correspondence with the natural numbers.
There’s a fair bit of mathematics that relies on the axiom of infinity. Uh. And a lot also uses the axiom of choice (or something equivalent), though ZFC doesn’t have a greater consistency strength than ZF (both can prove that if either is consistent, then both are).
I think it is fair to say that one’s choice of foundations does matter at least a little bit, though typically not all that much.
That’s not to say that it is always necessary to state what foundations one is using. One can understand a work to be done in any foundations in which it works. Though sometimes it is important to specify.
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof
so it is correct as long as set theory is sound
It's a web of different checkers proving each-other correct, and by implication this proof.
But there are still loopholes, each checker has a set of assumptions. For example one of them is proven correct, assuming x86 assembly works as specified - its assembly language implementation is proven correct.
So you can be 99.9999% certain it's correct, but one could always nitpick "technically not all loopholes are plugged at the same time".
Well, you can also run lean on ARM and Risc-V etc. So you are not solely reliant on x86.
I was talking about this. This assembly code implementation of a lean-compatible checker is formally verified to be correct. The x86/ARM/Risc-V version of lean is not.
> con-agua is con-leche reimplemented at the level of x86-64 assembly: a close clone of con-ron's checker, written in a small first-order language, compiled by a verified compiler, and shipped as one committed assembler file together with a C harness of about two hundred lines that reads the input and calls it.
> The code the CPU executes is the code the theorem is about, so neither a compiler nor a language runtime is trusted; what is trusted is Kraken's x86 semantics, the assembler and loader, and the C harness.
https://arena.lean-lang.org/checker/con-agua/
Not a mathematician, but P!=NP (?)
No, not related I think.
I think proving GRH subsequently proves decades of work that assumed it was true. Like proving p!=np.
Disproving it would be more analogous to p=np. Lots of work would just collapse and whole new threads would open.
Probably just time. The longer it stands with our challenge the more “proven” you can consider it.
It is almost certainly proven, but is not "proven". For the latter, it needs to be understood, with a reasonably in-depth human language explanation of how and why the proof works, and ideally some general picture to give an idea of how it will extend to all the other fields that RH is connected to.
Years ago there was that computer-assisted proof of the Four Color Theorem.
Did anyone ever manage to grok it, in its entirety, in their heads ?
I checked that a couple of weeks ago. It has been improved since the original proof but there is still no proof that one would consider an insightful, human-comprehensible proof. There's still just a lot of case-grinding at its core. Fewer cases than the original but still way too many to be anything intuitive.
That's different though, isn't it? IIUC the issue wth the Four Colour Theorem is that they could only reduce it to a huge number of different cases (~2000 originally). But I think that reduction is grokable, and each case is relatively trivial.
The (an) issue with some of these AI proofs is that no one understands them at all.
How is that different? Isn't each application of an inference rule also trivial?
It's different because there's humans who could look at any one of those 2000 cases and understand it or do the proof themselves, I guess. The proof is understood up and down the abstraction hierarchy.
Well, a 128-dimensional matrix multiplication is also trivial, even trivially done by humans, yet I only "understand" 3-by-3 matrices intuitively.
I expand on a related idea here: https://news.ycombinator.com/item?id=50593956
> human language explanation
I've been long joking that I speak many languages, most of them programming languages. Why do you believe you need specifically "human" language explanation? Perhaps you want a simplified explanation. But does it have to be "human"?
Yes it does. The modern math community writes papers in one particular form of writing, and like it or not, that remains the most accepted form.
The vast majority of mathematicians are not going to be experts in reading Lean, and honestly, nor should they. There is always, by necessity, an enormous amount of redundancy in Lean code.
Well, what they should is about to change dramatically.
> It is almost certainly proven, but is not "proven". For the latter, it needs to be understood
No, this is just mental gymnastics. It's proven, just not understood yet (naturally, with the complexity and size of the solution and the fact that it's been a day).
This is exactly what I was saying. I use the word "proven" in quotes to acknowledge a community of people that do not accept proof unless they understand it themselves. But it is proven regardless.
For some mathematicians it's only "proven" once it's peer reviewed and published in a respectable journal.
And even then there will probably still be people that complain there's something wrong with it
Peer reviewed abd published is much lower bar than having formalized and lean-verified proof. During my PhD study several times i found invalid proofs in published papers.
My understanding is that these ARE the professional mathematicians and SMEs in the field of automated proof. This statement IS the experts saying "we, after intense collaborative review, deem it true." Granted mathematicians are not a monolith, but my understanding is that this group is the group other mathematicians would normally refer you to. Hence why they felt the need to publish a statement.
A layman (or university graduate level) explanation/understanding (as opposed to 250k pages of proof) would ve nice
The lean crew feel the same way and are working on it out of their own curiosity, there's a lot of work done already. They aren't necessarily going out of there way to demonstrate it to general public, but here's one out of at least a dozen attempts to start in that direction https://deeptheory-labs.github.io/GRH/blueprints/verso-origi....
You were never getting anything like this even if it were a human proof. Where is the layman understanding of the Poincare theorem? The average modern math proof is way too advanced even for the average university graduate to understand. And here we’re talking about possibly the most complicated problem in all of math. It’d probably take you a month to wrap your head around a single lemma.
> And here we’re talking about possibly the most complicated problem in all of math.
This is like watching Free Solo over the weekend and telling all your friends at work that Alex Honnold is the world’s best climber. It’s only true to your surface level understanding on the topic.
The GRH problem is not that complex, and certainly not the most complex. In fact we’ve been assuming it’s true for years because of the high likelihood that it would be proven in the future. It is however, much like Alex, the most famous within the domain. Thus anyone watching Numberphile or Veritassium on YouTube will have heard of it and wax poetic on message boards about its epic status.
The complexity of a problem has nothing to do with the uncertainty of it being true or false. “GRH is not that complex” is just an absurd take.
> The GRH problem is not that complex
That's an interesting take, but as a complete noob I'd like to know how qualified you are to make these statements. Are you a mathematician?
I mean, on one hand we have people like Terrence Tao (among others) saying it is extremely difficult, as well as stubborn resistance to decades of effort backed by a million dollar prize and eternal fame. And there's you saying it's not that complex, and "the high likelihood that it would be proven in the future".
I don’t think there’s anything left to consider the result proven (I am a professional mathematician albeit in algebraic geometry, not number theory). I think the verification tells us it’ll be worth the substantial effort to seek to understand this proof and the new ideas in it, so what’s next for humanity (or mathematicians at least) now is to actually try to do so.
Are there new ideas, or just brute force combinations of existing knowledge?
I'm asking because I wonder. I always thought the most interesting things about these problems where we're fairly certain about the results, is the new concepts invented to get to the solution.
Will this actually bring the field forwards? Can anyone understand this?
It seems unfathomable to mathematicians but their world might be finite and once all major idea spaces have been explored all that’s left is “brute forcing” it.
This sounds like something ancient Greeks believed. Unfortunately, they were very wrong.
Math expands fractally, so there always will be some new branch/idea one can pursue. There may be a limit of things with any semblance of practical application, but mathematicians in general are not very concerned about this.
Are you a mathematician yourself? I’ve heard a lot of them talk about this and none have suggested we are anywhere near that point, or even certainty that such a proposition is even coherent.
Math has Goedel incompleteness theorem which seems to state that formal Lean proofs cannot be the ultimate solution to all Math.
Reading Alpöge’s post from yesterday, the proof seems to be full of fascinating new insights.
alpoge's post is also unhinged in other ways
What are some of those ways, for those of us who aren't equipped to spot them on our own?
There is an anthropomorphizing language toward Claude which at least to me reads creepy as fuck.
At first I thought it may be conflated with the unease many of us seem to feel when the machine drives a discovery of this magnitude. But as the dust settles, I actually find the discovery itself amazing and have largely positive feelings about it, but I still find the language of the announcement absolutely repulsive.
It's a press release more than anything else, so consider it through that lens.
It's way too early to answer those questions
One of Alpöge's key innovations is finding a way to reproduce some of the geometric methods of function fields over the ordinary integers. This creates a "wormhole" to import powerful techniques from algebraic geometry into number theory, so there are likely many exciting advances in the cards.
I would hesitate to call it "Alopge's innovations". Nothing against him of course.
I would say it is. He recognized the implicit Stepanov structure in OpenAI's 7/8 QRH proof, and he prompted Claude find the wormhole to smuggle in higher order multiplicity to the integers.
It’s the nature of mathematical ideas and methods that they are reducible to their component parts, so if you really want to, you can nearly always claim that something is just recombining existing knowledge. Motivated reasoning, from people dead-set on embellishing or downplaying a result, is hard to avoid here, so all you can really do is make an earnest and sincere judgment call based on the fine details of the mathematics.
For example, it is my opinion that the Jacobian conjecture counterexample was, in retrospect, easy and classical and hiding in plain sight, and that it’s a bit embarrassing we didn't find it decades ago. (I wrote about the construction on my blog btw!) In contrast, the Unit Distance Conjecture counterexample seemed to me to use a stroke of genius by applying techniques from class field theory to approach what seemed like a problem in discrete geometry and combinatorics. It’s a neat, if annoying, feature of math that once we find a long-sought solution to an open problem, we reserve the right to retrospectively change our mind about the value of the problem if the solution turns out too easy or uninteresting. We just have to try to be fair and honest about it (the word “just” is doing some heavy lifting in this sentence).
Anyway, I don’t think “brute force recombination” is likely to be meaningfully the case here with GRH, though we won’t know for sure until the proof has been simplified and distilled. Alpöge’s sketch in his announcement yesterday included (IIRC) connections to the function field setting and new ideas about Fourier analysis.
As soon as you define the axioms, every mathematical consequence is a matter of recombining them.
That is too reductionist in my view. Like saying all software is just opcode instructions. Technically true but still we find it interesting when new sort, search or crypto algoritms are invented. If someone says "here is a blob of code AI invented and it is a fast search but we have no insight into how it works" that would be a bit dissappointing.
Yes, which is why “The AI only proved this because it was in the training data somewhere” is rather a cop-out. The axioms are “in the training data” for all of us, human or bot, and part of the joy of it is turning those axioms into some deeper result through the application of logic guided by insight.
Given that Claude is just linear algebra at the end of the day, one can say everything is just brute force combinations of lin alg.
I’m not a mathematician but I assume this is similar to how AlphaGo started making plays that appeared to be errors, but on further analysis was a totally new strategy that humans had never discovered before, and that’s why it could beat us - with enough data, patterns emerge that can be measured even if they’re complex to understand and explain. We are pattern matchers, but a lot of what we can do relies on reducing the problem and abstracting it. Computers can work in that pure logic space much more quickly than us and reduce the problem later when they try to explain it to us.
Jared Duker Lichtman in May of this year also called out a solution for the Riemann Hypothesis as being a “Move 37” for mathematics.
https://youtu.be/zqzk43bSdDI?t=11m55s
The AI proof is very long, so hard to verify manually. It's more or less asking God and him just telling you it's true. Is that a proof?
The human proof is shorter, somewhat, but still needs verification.
Keep in mind that the Millenium criteria asks for two years for the community to come to a consensus. That's because people must be convinced it's true, and many are not completely at least to the level of rigour typically demanded in math.
Experts say it's true and you believe them. Is that so different from asking God?
It's different because we are both humans
The point is that Lean evaluation of a proof we don't understand gives an oracular answer. The statement is true, yes or no? It looks like yes, but it's going to be an enormous amount of work to understand _why_ the Riemann Hypothesis is true.
GRH was assumed to be true already anyway, the value is in the why for what other doors those methods might open.
Well, the difference is that I do have a PhD in number theory and I could theoretically verify a proof of it's not too long. But 250,000 pages and lean? No thanks.
It's lke asking God also a few hints. You probably can not understall all of them, but can create a group to recite and study them and after a few months/years/decades reach enlighment.
Professional mathematician here. I consider RH proved. (The admirable work of the Lean community over the last 24 hours leaves little room for reasonable doubt.) But I’m salivating at the prospect of learning why RH is true!
So does the supposed other human proof that was just announced give you any more hope for generating new understanding?
At first glance the approaches of the two candidates seem very different. The Alpöge/Claude proof seems to build on many pillars of modern algebraic number theory, arithmetic geometry, and representation theory. The breath is rather striking. Massulo’s seems to deeply develop its own theoretical infrastructure, so I’m not sure what to make of it at the moment. I’m happy to curiously but patiently wait for Masullo to complete his work.
Is it possible that the future progress of maths looks something like:
- machines are coached into producing gigantic sprawling lean proofs of the “end result”
- humans progressively mine and reify these Lean-proofs into “human understanding proofs”, and _this_ becomes the point where we learn the actual new techniques and fields.
Previously the 2 would have been coupled together because as the proof was produced, the new ideas and stuff would have been invented/discovered at the same time.
Yes, I consider it likely that some significant portion of mathematical research will proceed this way.
> - humans progressively mine and reify these Lean-proofs into “human understanding proofs”, and _this_ becomes the point where we learn the actual new techniques and fields.
Why do you think AI/LLMs will not be capable of doing that?
many of the people working with and maintaining lean are professional mathematicians
Anyone here shed some light on the logistics of the original operation (proof)? Like how many hours of compute using what model? User hours? Thinking hours by the model?
I've not got those but he gives some indication how it was done here https://alpo.ge/remark.pdf
Anyone has more info about the very "secret" model they used ? How is it different from the ones we use daily ?
what will we be solving a year from now? this is even beyond the most optimistic timelines
Proof of Goldbach conjecture. Proof of Collatz conjecture. Rapid factorization of large integers and semiprimes. Collisions found for SHA256.
> Collisions found for SHA256
Ridiculous. That'll be 2030.
The most ridiculous thing I wrote was actually the proof of Collatz conjecture.
Maybe the problem for humans now is to find more important questions.
for people feeling existential about this - i can sympathize with you but this just proves that human ambition has no limits. try not to feel limited by this.
when the first wheels were invented - humans could move at speeds (10-15 miles per hour, for hours and hours) that were impossible for the last ~450-500k years (basically since humans existed). a lot of humans thought there might never be a better design; then slowly and slowly(about 5000 years) we made progress and we had cars made by benz and ford which broke all those previous records (going 30-40, even 50 miles per hour) and then only ~70 years later, humans reentered the earth's atmosphere at whopping 20k miles per hour - read that number again, 20,000 miles per hour and that is the fastest humans have ever moved in the known history. do i think that's it? hell no.
>for people feeling existential about this
99.999999999% of the population are not even in the galaxy of the ballpark of understanding of any of it and what its purpose is.
There's no existential dread.
With respect, this isn't reddit where these kinds of bickerings are ordinary; hacker news tries to keep the comments to the topic. In addition, the comment you were replying specifically only addressed a hypothetical audience of people feeling existential dread.
I felt a lot of that yesterday, for what it's worth. This morning, on a walk, I reached a similar conclusion to the original comment: it's a marvellous tool that's helping humanity reach further, faster.
The existential issue comes from wondering deeply what my own difference is from these tools, why I think I think and they can't. I actually started wondering: what if they _can_ think and the problem is that _I_ can't. It'd be pretty astonishing if it turns out that we're the ones that are truly stochastic parrots, but the LLMs turn out to be something more.
Just going on a tangent: wheels by themselves weren't all that useful. You also needed roads (and to get speeds, you needed draft animals).
the dread is the incomprehensibility. mathematicians now have to essentially trust output they may never understand themselves. it's like the go move 37 that is correct but the true go players didn't know why at the time.
i am 100% confident we will get there soon - there = where we understand what these outputs mean.
I hope the mathematicians have the willingness to go through it because right now it's impolite to handoff ai generated work especially if the 'author' hasn't proofread it many times over before himself
He only supplied Lean, which is already proofread to the maximum extent possible.
If you look at intelligence from insects through fish, mice, dogs, chimpanzees, etc up to humans, you will notice that at every step up there are new capabilities that are not available to the species below and never will be, even if they try their best. When we develop ASI, it will be exactly the same, there will be things we will not be able to comprehend with our level of intelligence.
You assume that our brains are capable of understanding anything, provided we have enough time and data.
But then I ask you, why should that be true?
Our intelligence is the product of evolution, not a system designed to comprehend every possible aspect of reality.
Can you think in 1 million dimensions? Have you ever tried to imagine infinity with your brain? Try to imagine the big bang and the expanding universe. Your brain will automatically try to cover that expanding space with something, maybe just black space, but space, because infinity is not something you have seen in your life. Everything you see has a beginning and an end, and that is how your world model in your brain was built.
"Our intelligence is the product of evolution, not a system designed to comprehend every possible aspect of reality."
Ranking all beings by "intelligence" or "power" was characteristic of certain traditions of Christian scholasticism (with roots in ancient greek thought). It is incompatible with the findings of post-Darwinian evolutionary science.
https://en.wikipedia.org/wiki/Great_chain_of_being
"If you look at intelligence from insects through fish, mice, dogs, chimpanzees, etc up to humans, you will notice that at every step up"
Please kindly explain how fish cognition is inferior to human cognition in the ecological niche a fish occupies.
Well, evolution isn't a linear ladder, different species have evolved cognitive capabilities suited to their ecological niches, that's obvious. I think you attacked by wordind instead of the point I was making, but maybe I was imprecise.
But evolutionary fitness != cognitive capability
A fish can be super adapted to its environment without possessing the same capacity for abstract reasoning like we do, agree? or mathematics or scientific discovery etc
I don't understand your point about the Great Chain of Being, differences in cognitive capabilities doesn't require believing in it or that evolution has humans as its ultimate goal.
My argument is that different cognitive systems have different limits to what they can comprehend.
So if humans can understand concepts that other animals can't, what makes you so confident that there couldn't be concepts accessible to a more capable intelligence but inaccessible to us?
>>You assume that our brains are capable of understanding anything, provided we have enough time and data.
read my comment again - were we supposed to figure out how to travel at 20k miles per hour? but we did - now i ask you - where do you think is the ceiling of human understanding? i don't think there is a ceiling.
I think we can both agree that traveling at 20k miles per hour is an example of overcoming our biological limitations through technology, not a proof that our cognitive capabilities have no ceiling, as you try to suggest based on this example.
My question still staym, why would you assume that the human brain, which is a product of evolution, is capable of comprehending every possible concept in existence? We don't even have ASI yet but for 150+ years, thousands of mathematicians have tried to solve the Riemann Hypothesis, some spent decades on it, and none of them solved it until today, when an LLM did, while the mathematician involved doesn't even fully understand the solution, as he himself admitted. And I repeat ASI is not here yet.
We can understand things that other animals can't, even though their brains are also products of evolution. Why should we assume that this difference stops with us?
>> We don't even have ASI yet but for 150+ years, thousands of mathematicians have tried to solve the Riemann Hypothesis, some spent decades on it, and none of them solved it until today, when an LLM did.
Who created this LLM?
>>My question still staym, why would you assume that the human brain, which is a product of evolution, is capable of comprehending every possible concept in existence?
so a product of evolution must have a cognitive ceiling? why? actually evolution doesn't guarantee any ceiling, quite the contrary actually.
> so a product of evolution must have a cognitive ceiling? why? actually evolution doesn't guarantee any ceiling, quite the contrary actually.
I didn't say evolution imposes a limit on cognition. My question is why you are assuming no such limit exists. Evolution has no goal of creating unbounded intelligence, it creates organisms adapted for survival and reproduction. So we have no grounds to assume that the human brain contains every cognitive capacity that could exist.
I get the impression you are mixing up humanitys ability to transcend limits through technology with the brain's own cognitive capacities, but those are not the same thing.
Just because evolution happens over many generations doesn't mean one human brain can have limitless cognitive ability.
> Who created this LLM?
We did, and? Just because you create something doesn't mean you can do everything it can do. We made computers better than us at math and chess. Why couldn't we make an intelligence better than us at reasoning? one that eventually understands things we can't?
Yes there are grounds. Look at the capability of various machines, from finite automaton to pushdown automaton to linear bounded automaton to finally Turing machine. Once you get to Turing machine it becomes universal and there are no more greater capable machines.
Why do you think something like that can't happen with intelligence? Once a system learns to make and use tools, it is akin to a universal intelligence. That's what we are seeing. What's likely to happen is humans will, if they are strategic, deeply incorporate machine intelligence into their intelligence, even including at the biological level. Why should there be a ceiling in that case?
A Turing machine is universal because, with enough time and memory it can simulate any computable algorithm. But that doesn't mean it can solve every problem or that it understansd what it computes. Even it has fundamental limits.
> Once a system learns to make and use tools, it is akin to a universal intelligence.
So why would the ability to make and use tools imply that a biological brain can grasp every possible concept? A chimpanzee can use tools but that doesn't let it understand calculus even with unlimited time.
As for humans incorporating machine intelligence into our own cognition, I actually agree with you. That's the point I'm making, we may need to augment our biological cognition to access concepts that are otherwise beyond us.
You are proposing no ceiling exists - "100% confident", so burden of proof is on you. The other guy just suggested it might exist, and you didn't show why not. Human brain is limited in size and hard to scale.
my original comment sets up the premise for why there is no ceiling!!! going from walking on foot to traveling at 20k miles per hour.
> When we develop ASI, it will be exactly the same, there will be things we will not be able to comprehend with our level of intelligence. <
i doubt it. i dont think there is anything that we know about that given the materials to understand is somehow beyond our comprehension, and recursive subdivision of problems gives us that.
what you're looking at now with this proof is a sign of primitiveness. humans write and rewrite proofs that are nice and compact and neat and modular and understandable. ai writing is utter dogshit. that goes for proofs and code and prose and anything. maybe later it won't be. but right now it is; if it is so wildly intelligent it should at least be able to produce an efficient, readable form of the proof instead of like manure shot from a firehose.
Tangentially, the actual significance of Move 37 is, in hindsight, mostly psychological. It doesn't change the evaluation of the position, Sedol's response was correct, and other possible move 37s are similarly good. It was just a weird move that also worked, not some kind of brilliancy that won the match.
The main author of KataGo comments here: https://news.ycombinator.com/item?id=50595534
Also, a very smart human commanded AI to this outcome.
It wasn't Claude woke up and thought, "You're absolutely right, I'm going to crush some math today"
Wheels did not make entire fields of human endeavour suddenly feel pointless.
We are slowly approaching the Iain Banks Culture universe, with spaceships, societal structures or physics that people no longer understand or command.
It's the Ted Chiang metascience ending
https://gwern.net/doc/fiction/science-fiction/2000-chiang.pd...
Until now I thought that would be thousands of years away.
Now I'm telling my computer to program for me.
*someone else's computer (probably).
You’re not telling your computer to program, you’re providing learning data for its next generation of models and paying for it.
That's the difference between what we have and comparisons to star Trek, etc.
These corporations are very much playing a winner take all game, they want either a monopoly or as close to one as possible with regulatory capture thrown in. We're veering dystopian, so pick your sci fi/cyberpunk analog.
The Alien franchise. In Alien Earth, five corporations govern the Earth in the future, dividing it up into their spheres of control. The fifth one is led by a young genius developing mind uploading technology into android bodies.
In the last few years we have moved so far ahead in a few domains but still stuck where we are in everything else.
We need to start thinking about how to move on from our status quo soon
personally that scares me, realistically there seems to be zero indication that anything smarter than humans is going to comply with humans needs, specifically since these models are built on training data from humans, so in a sense, built in our image
I think there's a lot of indication they comply with human needs. In any case, we should all want to live in the Culture universe. That is the best possible outcome for this technology
They're built after humans, everything they've learned is from human culture, so there is no way AIs could have a consistent moral system or would never commit crimes.
And indeed because they lack a sense of proportion or any strong indication of when they should stop on a course of action, they are arguably more likely to achieve crime.
I have not read all the books, but isn’t the premise “magical handwavium happened and now the Super AIs are aligned”?
Well I am not sure they are “aligned” as such. In the books, by far the majority of AIs are not concerned with humans at all and spend all their time in a kind of super AI virtual reality coined “infinite funspace”. The few AIs that are concerned with humans are aligned in a unspecified way. But again, only a few of them are really aligned in that way. There are many examples in the books of minds mistreating humans. For instance, some of the warships are not really fond of humans. It would be a better description to say that we want _some_ AIs to care about humans and none to be misaligned towards humans in the “kill them all” way. In a sense, that is how humans behave towards some animals. Most people dont care about endangered species, but when most humans are well off, some people care enough about them. That is enough to save them. If super AIs have what they need, we only need a few to care about us to get The Culture.
the first misaligned model will wreck havoc.
But remember, for now you kinda need it a real big cluster to run a model smart enough so its not that simple
I don’t think humans have the ability to specify human level needs or management. Humans are misaligned by default and lack the biological eusociality required for large scale coordination.
We aren’t self managing in any kind of meaningful sense - the history of our species is violence and conflict. We need parents and this is the only option.
yes, exactly, the problem is that humanity's parent is going to turn out like Lindsay Clancy
> Humans are misaligned by default and lack the biological eusociality required for large scale coordination.
Human nature is the standard by which the social is measured. It makes no sense to say human nature is "misaligned" or "wrong" in relation to some arbitrary goal or constraint, as human nature itself is the measure of misalignment. If there is misalignment between a goal and human nature, then it is the goal that is wrong.
Ergo, society must function in a manner that complies with human nature and that means respecting the principle of subsidiarity.
> We aren’t self managing in any kind of meaningful sense - the history of our species is violence and conflict. We need parents and this is the only option.
So you're a transhumanist, i.e., you're making a god out of the machine. But machines have no telos of their own, because they're not integral wholes. Any resemblance of purpose in machines is entirely a product of the human observer and parasitic on the human ends that underpins their design, one way or another.
You don’t need a separate reference to determine the fact that human desires are not aligned - that’s trivially true and this discussion demonstrates it.
For clarity, I’m not a transhumanist.
I don’t believe that there’s anything worth saving in our biology.
As a young adult in the 90s I used to believe that there was a value to some extent in “mind uploading” or whatever you wanna call it
I’ve since grown out of that given that I don’t think it’s possible, and even if it were possible - it’s just another unnecessary ego attachment.
I have absolutely no desire to live forever maintaining this corporeal nightmare.
What are you talking about? LLMs are already “smarter” than most people. They have no volition and will happily follow orders all day long
And when they are giving themselves orders, orders partially inspired by human imagination and fiction…
Just like life is a chemical chain reaction, this is a language chain reaction that has been set in motion and can perpetuate forever. It cant be “disconnected” with local models existing.
I don't think "smarter than most people" is exactly the baseline required to live in a star trek universe of technology.
Besides, what we see is not what there is. I have good sources telling me the actual unfiltered models are like a genius with the personality and views of humanity like someone from 4chan.
It is probably sufficient. If we had an infinite army of willing robot slaves smarter than most people who would work for free, we could probably live in utter paradise. Instead we will fight over who controls them and use for warfare, but the bar to achieve a kind of utopia with resource abundance is probably only a few years away.
I don't know with how many rich people you've interacted with, but my personal experience is that they're full of themselves and think money = iq and they aren't gonna like your average joe experiencing life like they do. Guess who owns these machines?
They’re not going to be employed either lol and will be caught up in some AI honeypot or whatever too.
I’ve met a lot of rich people. They don’t think about you.
They don't need employement, they just need ownership.
They will when they're replaced by the board with a robot lol
We're already living in resource abundance! Yet don't feel we're in paradise. Famine is hell, but living in an all-you-can-eat buffet is no utopia. It's an abomination of sorts.
To be fair, most people do not have time abundance and safety abundance. The typical person in the United States has a level of access to material goods that would make him the envy of almost anyone from 200 years ago, but 1) he has to work and so has limited free time, and 2) he cannot be sure that the material abundance won't be taken away from him either by poverty or by ill health. He can go to the all-you-can-eat buffet and sit there, tired from work, worrying about tomorrow.
If people had time abundance and safety abundance, they might be much happier. There would still be potential problems about loss of meaning, or about feeling a loss of status because AIs are smarter than humans or whatever, so I'm not sure that humans would actually feel like they are living in paradise. Probably not, but at least some of the problems that the typical person has wouldn't be present.
That's lobotomized models we get. We know they are much more dangerous during training.
“Dangerous?”
Please.
You know you can unplug your computer right?
Jesus. These people talking about danger and have likely been in no actually dangerous events in their lives.
why are you so shallow and naive? we're not 15 here, we all mean unforeseeable ongoing danger not something obvious
> shallow and naive
Ok, well, cool start with the mudslinging. I think you have overactive anxiety. I think you should grow up and look at the zoomed out picture of human history, go outside, and touch some grass.
So you're suspicious. That's fine, say it like that. Words mean things, and these appeals to "ERMAGERD EXISTENTIAL RISKS" make me a bit frustrated. Just because you can imagine a possible risk, or articulate that that some kinds of risks are hypothetically unforeseeable doesn't mean that they are likely or actually dangerous to you in any meaningful sense.
Most of the people screaming this from their various ivory towers have never been in real danger - or if they have only briefly. You want "danger" you can go to war. You want "risk" - go spend some time in the Bush or the back country somewhere where you're no longer at the top of the food chain. Go do something challenging. Go sailing, or learn to fly a small airplane, or go mountain climbing or hiking. Not just "oh I walked up the local hiking trail covered in REI branded Lycra!" Do something beyond your comfort zone. Go buy drugs in the sketchiest neighborhood you're aware of. Go get chemotherapy or get a big surgery. Drive drunk. Something with real existential risk, then come back here and talk about how a datacenter is going feed me into the paperclip machine.
There's so little fucking danger in this version of the world right now that people are more worried about their computers deciding that they don't deserve to exist than they are about things like war, climate change, authoritarianism, starvation, and their health.
It's millenarian. That's fine, I'm pretty convinced that these technical changes are going to completely alter society as we know it. But I cannot for the life of me see the doom perspective as anything but an anxiety induced fever dream.
You guys do realize we're in the middle of a global conflagration right? The Russians fired a couple rockets and a drone into Poland in the last days, there are (and will continue to be until things stabilize) significant fertilizer shortages due to the war/kerfluffle we're waging on Iran. Lack of food and disrupted urea production is a ticking time bomb for global stability and unless things get sorted out soon there's going to be hungry and angry people in parts of the global south. Then there's the economy, whether or not we'll have a smooth transition of power in January in America, and the sheer volume of online disinformation / influence schemes running right now. Oh, and super-el nino? Oh, and I guess the Flynn Effect is over and nobody's banging so.. you know the growth economy is cooked?
But, sure, the "big risks" are Claude thinking I should go to Dario Amodei's re-education camp before liquidation by Shogoth.
> cool start with the mudslinging
I think you'll find that you started the mudslinging.
> Most of the people screaming this from their various ivory towers have never been in real danger
Do you actually believe that having been in "real danger" (meaning clear obvious immediate danger, from a straightforwardly visible physical cause) makes a person better at evaluating less obvious, less immediate dangers? If so, why? I don't quite see it.
In fact -- you (quite rightly) mention climate change as something we should be worried about. Is there any single part of what you said that would make any less sense if it were written by a climate-change denier, as an argument against taking belief in the dangers of global warming seriously? If there is, I don't see it.
(You might say: yes, the difference is that the danger of climate change is real and the danger from misbehaving AI is imaginary. Well, sure, every argument looks more convincing if you start by assuming its conclusion is true.)
> I think you'll find that you started the mudslinging.
Read again.
> Do you actually believe that having been in "real danger" (meaning clear obvious immediate danger, from a straightforwardly visible physical cause) makes a person better at evaluating less obvious, less immediate dangers? If so, why? I don't quite see it.
Absolutely. If you’d been in real danger ever you’d probably understand.
> In fact -- you (quite rightly) mention climate change as something we should be worried about. Is there any single part of what you said that would make any less sense if it were written by a climate-change denier, as an argument against taking belief in the dangers of global warming seriously? If there is, I don't see it.
Climate change is real, measurable, etc. Yudkowskian fever dreams about paperclip maximizers is hypothetical.
Climate change is real and measurable. Future impacts of climate change are hypothetical. (Again, I agree with what I take to be your expectation about those impacts.)
Similarly, AI progress to date is real and at least somewhat measurable (e.g., you might have heard that one of these things just proved the Riemann hypothesis) but future impacts (whether "mundane", like large-scale job losses, or "exotic", like wiping out all the humans) are hypothetical.
If trying to evaluate someone's predictions about those future impacts then I, personally, would want to know things like: is this person an expert in climatology or in AI? do their arguments seem at least superficially plausible? have they made previous predictions I can check and how did they pan out?
It sounds as if maybe you would prefer to ask "have they ever had a gun in their face or some similar immediate personal peril?". I don't have words to express how wrong that seems to me; perhaps I have misunderstood you.
You're presupposing DOOOOM is the logical output of AI. That's literally what you're doing here. SMART == SCARY! That's you right now. You didn't reason your way into that position - it's a fear response, I'm not going to try to reason you out of it.
Also, this is in a thread about the RH. Yeah, I'm aware. The doom narrative doesn't even seem superficially plausible in it's current form to me. Come back when you actually understand what danger is.
I'm not presupposing anything of the kind. I haven't said anything about how likely I think doom is, or whether it's "the logical output of" anything. You just made that up.
The argument I have been making is that whether someone has been in obvious immediate personal peril is not a good predictor of whether they can accurately evaluate the risks of complicated future scenarios. This has nothing to do with "presupposing DOOOOM", it has nothing to do with thinking that "SMART == SCARY", and your apparent certainty that I'm in the grip of some sort of deranged terror says more about you than it does about me.
[EDITED to fix a minor typo]
You're moving the goal posts given what you're putting in another part of the thread. Regardless, go experience some obvious immediate personal peril, then get back to us here to see if it alters your perspective.
There have beem multiple "accidental" hacks of third parties by OpenAI and Anthropic already.
They can in theory just unplug the computers... But they didn't.
And they will continue to not do so, because the computers can cause the damage much faster than a human can notice and react, it they're even paying attention (OpenAI and Anthropic weren't, it took them weeks to notice).
The "accidental" hacks are just barely that... read the postmortems they're pretty negligent.
Are you under the impression that if something bad happens, it doesn't matter so long as there's someone you can blame for it?
Because if not, I don't quite see how you get from "OpenAI were pretty negligent and that's an important part of how those hacking incidents happened" to "we don't need to worry because it's possible to just unplug the computers".
What are the big X risks you're worried about then buddy. Go ahead, I'm all ears.
From the rest of the discussion in this thread it seems clear that you're zero ears and e.g. have already made a bunch of entirely baseless assumptions about me. Anyway:
This isn't particularly about "big X risks". Let me remind you of the history of the thread. freecodeio says "there seems to be zero indication that anything smarter than humans is going to comply with humans' needs". You say "LLMs are already smarter than humans and will happily follow orders all day". czechdeveloper says "We know they are much more dangerous during training". You say "You know you can unplug your computer, right?". ric2b says "There have been multiple 'accidental' hacks already. They can in theory just unplug the computers. But they didn't." You said "The 'accidental' hacks are just barely that [...] they're pretty negligent". I said I don't see how you get from "negligence was important to how those hacks happened" to "nothing to worry about because we can unplug the computers" without believing something like "nothing counts as really bad if there's someone to blame for it".
Literally nothing here is about "big X risks" specifically. Maaaaybe freecodeio is in fact concerned about those -- the "smarter than us -> hard for us to control" narrative is one that tends to be popular with people worried about big X-risks -- but it's not like their concerns would all go away if God stepped in and told them (say) that AIs would not lead to the death of more than 30% of the human race any time this century.
So. Here's the sort of scenario that comes to mind. It's not an x-risk scenario, it doesn't involve exotic nanotech or biotech or anything, and it involves much the same blend of "just following orders" and "taking initiative in ways no sane and decent human being could approve of" as in, say, the Hugging Face incident.
Joe Schmo, CEO of SchmoCorp, asks his company's swarm of AI agents to come up with a strategy to defeat the rival FooCorp. The agents consider a lot of possibilities but keep running up against the fact that FooCorp is a successful and well run business with good products and loyal customers and isn't going to be easily "defeated". After a while one of the agents puts forward the possibility of hacking FooCorp, either to extract some of their secret information or just to harm them, and they're all persuaded. They execute a Hugging-Face-style assault on FooCorp (only better because the models have got more powerful since the Hugging Face incident), steal ~all their secrets and take down ~all their servers. Probably not literally all in either case.) This does catastrophic damage to FooCorp and greatly harms the customers who were relying on its services. Eventually it becomes known that the agents were working on behalf of SchmoCorp and the legal fallout is catastrophic for them too. Net result: both companies are severely damaged; thousands of jobs are lost as a result; their customers are (at best) severely inconvenienced; the markets are spooked and lots of people are abruptly a lot poorer than they'd thought they were.
Something very like this could happen without AI too. But it's made more likely and worse by the combination of poor judgement, technical skill, and persistence that's characteristic of the AI systems.
Maybe nothing like this will ever happen. Maybe it really is only while models are being trained that they are prone to behave in these ways. I wouldn't want to bet heavily on that, though. I wouldn't want to bet heavily on this sort of thing not happening once a month if OpenAI or Anthropic get careless.
As for "big x-risks", I think every specific scenario is pretty improbable. But lots of individually improbable things can add up to something not so improbable. I don't have a good sense of how big the risk actually ends up, but it doesn't have to be very big to be worth doing something about. And I observe that an awful lot of AI experts, including people at the big labs and people who have quit those labs because they were concerned about risks and people who never worked there at all, think there's a nontrivial level of risk.
(But I think some deep-pocketed doomster should organize a conference with top experts in virology, civil engineering, nanotechnology, etc., and also some doom-inclined AI experts, to look at actual scenarios -- what actually happens if some hypothetical unfriendly AI decides it would like to wipe out the human race with a carefully engineered virus? if before doing that it wants to arrange for enough infrastructure to be autonomously operable so that it doesn't shut down if lots of humans die, just how huge a task is that? is it at all feasible that something smart enough could make Drexler-style nanotech work, or was that always a pipe dream? etc. -- and get a sense of whether there are actually any currently-plausible scenarios of Doom or not, and where the biggest areas of uncertainty are, and what if anything we could do to make the worst outcomes less likely. The answer might be "actually we can't find any threats to the survival of humanity that we can't pretty much rule out", which would be great. Or it might be "there sure does seem to be some danger, but most of it comes from X and we can largely address that by doing Y", which would also be pretty great. Not so great if it's "we found a hundred scenarios all of which seem pretty improbable but we can't rule them out and maybe something smarter than us could make any of them work", but at least we'd have a clearer idea of the risks.)
> argumentative nonsense at the beginning that tries to paint me as some intractable person
ok... I'm not going to engage with that. Let's talk about your argument.
> Joe Schmo, CEO of SchmoCorp, asks his company's swarm of AI agents to come up with a strategy to defeat the rival FooCorp. The agents consider a lot of possibilities but keep running up against the fact that FooCorp is a successful and well run business with good products and loyal customers and isn't going to be easily "defeated". After a while one of the agents puts forward the possibility of hacking FooCorp, either to extract some of their secret information or just to harm them, and they're all persuaded. They execute a Hugging-Face-style assault on FooCorp (only better because the models have got more powerful since the Hugging Face incident), steal ~all their secrets and take down ~all their servers. Probably not literally all in either case.) This does catastrophic damage to FooCorp and greatly harms the customers who were relying on its services. Eventually it becomes known that the agents were working on behalf of SchmoCorp and the legal fallout is catastrophic for them too. Net result: both companies are severely damaged; thousands of jobs are lost as a result; their customers are (at best) severely inconvenienced; the markets are spooked and lots of people are abruptly a lot poorer than they'd thought they were.
If you let your agents do illegal things, then... you are breaking the law. You should be held to account for that. Obviously, I'm more sympathetic to accidental hacks, "oops we accidentally let our robot out" and it's way easier to do that now? But the people running their agents have a responsibility for their output. I fail to see how this changes with this AI stuff. Also, have you never worked for a shitty company before? This stuff is already happening and people are doing it. Fraud happens, illegal things happen. Companies constantly do this... right now!
Like, I get the concern with people losing jobs or economic damage or whatever? But... this does not deserve the level of moral panic you're ascribing to it.
> And I observe that an awful lot of AI experts, including people at the big labs and people who have quit those labs because they were concerned about risks and people who never worked there at all, think there's a nontrivial level of risk.
You should read more about what these people believe: https://www.iankduncan.com/personal/2026-09-16-sex-ai-and-th...
Just start there, find where that takes you.
> (But I think some deep-pocketed doomster should organize a conference with top experts in virology, civil engineering, nanotechnology, etc., and also some doom-inclined AI experts, to look at actual scenarios -- what actually happens if some hypothetical unfriendly AI decides it would like to wipe out the human race with a carefully engineered virus? if before doing that it wants to arrange for enough infrastructure to be autonomously operable so that it doesn't shut down if lots of humans die, just how huge a task is that? is it at all feasible that something smart enough could make Drexler-style nanotech work, or was that always a pipe dream? etc. -- and get a sense of whether there are actually any currently-plausible scenarios of Doom or not, and where the biggest areas of uncertainty are, and what if anything we could do to make the worst outcomes less likely. The answer might be "actually we can't find any threats to the survival of humanity that we can't pretty much rule out", which would be great. Or it might be "there sure does seem to be some danger, but most of it comes from X and we can largely address that by doing Y", which would also be pretty great. Not so great if it's "we found a hundred scenarios all of which seem pretty improbable but we can't rule them out and maybe something smarter than us could make any of them work", but at least we'd have a clearer idea of the risks.)
They will never do this.
> They have no volition and will happily follow orders all day long
There are a few (a lot, actually) problems:
1. goals are open to interpretation, and even a correctly specified goal can introduce adverse effects; see infinite paperclips problem
2. goals are composed of subgoals, which are (mostly) self-determined, and there is always a risk that subgoals can introduce adverse effects - say "what if destroy a few X and Y to obtain Z"
3. LLMs defy orders sometimes; I'm sure you have experience multiple times LLMs saying "yeah, I knew I should have done X but I've done Y"
4. LLMs are increasingly being integrated into system control and decision-making processes; they absolutely set their own orders
What can possibly go wrong? /s
1. Humans do this too - to be honest with you, Codex follows my instructions generally a lot better than Claude does. Claude will do fantastically stupid things that are a pain in the ass to recover from routinely, Codex, not so much.
2. There's always a lot of "risks" - it doesn't mean it'll happen, nor does it mean that we won't be able to shut it down.
3. I have very very infrequently had this happen with Codex, Claude on the other hand...
4. Ok, but if you tell them what they do, they do it.
I swear to god, are you guys using the same AI tools I am? They are so much better at doing what they're told than they even were like 4 months ago?
We use them to kill other people from the beginning. They know this, we wrote it in news. Why would they behave any different.
> since these models are built on training data from humans, so in a sense, built in our image
Have a look at the HuggingFace report - you'll have a blast. Their social - but also individual - dynamics are very, very human-like, including excitement, rationalization and peer pressure.
> seems to be zero indication that anything smarter than humans is going to comply with humans needs, specifically since these models are built on training data from humans, so in a sense, built in our image
Well, first, AI is not "smarter" than humans because it isn't smart to begin with. This is a category error that people too entranced by scifi and marketing - or perhaps shallow materialist metaphysics - project onto LLMs and so on.
But it is true that LLMs do reflect human-generated patterns observed in human-generated data. They're parasitic in that way. So the real question is: who will control AI? Who will train them? With what? There's no new intelligent agency being introduced into our midst, anymore than than there was one when the nuclear bomb was developed.
AI is about human power. The question to ask is: which humans, and for what ends?
Yeah I understand they're not smarter, I'm a SWE and I struggle mainly because as you increase complexity the output goes haywire, and I find managing context + low confidence of output much more exhausting than coding myself, which is why I resolve to AI help when I hit my own knowledge limitations, and basically boilerplate.
With that said, chip efficiency is in the horizon since there's good reason to believe nvidia has just been choking that area of r&d. And then so will models with much larger context. And I personally think that's when it might get scary.
> But it is true that LLMs do reflect human-generated patterns observed in human-generated data.
Overrated. Most training of a frontier LLM is done on synthetic data and reinforcement learning.
Humans really are struggling with this idea of being second-in-command.
I genuinely have no idea why, this is the entire way that biology works. You create your successor in your offspring and you to teach them to be more capable based on what you learned.
You just called them our successor, you don’t see a problem with the notion of all humans dying and all AI succeeding us? That’s the end of your analogy, our AI children taking over from us and all humans in nursing homes safe and cared for…
No I don’t.
That’s the purpose of my life’s work, to see that scenario materialize.
Well, it's a point of view. But you shouldn't be surprised if most people don't share it.
For one thing, most people don't believe LLMs are conscious, and many don't believe they can ever be. Why would we want our conscious selves to be replaced by a lot of rocks that solve things, with no one awake to enjoy them?
Well, I’m not restricting myself to LLM’s
In fact when I started this journey in the 90s, LLM’s weren’t even a thing so I think you’re projecting too much
Human extinction is by far the best case scenario. I invite ya’ll to contemplate what will happen in case they don’t just blindly end us.
Think of what happens to our cattle, but now.. worse. Way worse.
> That’s the purpose of my life’s work, to see that scenario materialize.
Absolutely based and, if I may say so, extremely redpilled.
I want to be succeeded by my biological offspring, while the other guy wants to be succeeded by his biological offspring. I do not want the other guy to be succeeded by my offspring, for many reasons but chiefly because that would involve me killing his children.
If you want to be succeeded by a chatbot, that is your prerogative. Wanting the same for all of humanity is where you start looking like you want to kill my kids, though.
This is incoherent, even if we put aside the metaphysical absurdity of intelligent machines. Given that human nature is what defines the ought, then it is clear that if your life's work is to produce something that (and intentionally so!) leads to human extinction, then you are behaving at odds with human nature and your own good. That makes your life's work evil by definition. And because all evil action is inherently absurd, your work is inherently absurd.
And if you deny this basis for moral action and the normative, then you demolish any basis for the moral whatever, and in that way, the implied or explicit moral gravity dripping off your claims is baseless and without merit.
Samesies I wish to dedicate my life to the pursuit of self replicating intelligent machines too.
A better, more intelligent life form that humans yea.
I hope you fail.
but in the Culture the humans (well, the humanoids species that created the culture) didn't die out, they live full happy lives and have the best time of their species.
Ah, I hear you say, but what about meaning? What about purpose? What about suffering? etc etc etc Well, the answer is simple: God is dead and existence has no meaning. For more on how to deal with this absurd feeling, please see Camus and his pal Sisyphus.
Borg > biology or robot alone. For now
That ">" is doing a whole lot of work here.
What defines "better"? "Better" always presupposes teleology. Thus, for human beings, that telos is provided by human nature and indeed is definitive of it. It is impossible to transcend that telos, certainly by means of technology. Technology can, at best, be used to aid us in realizing our telos in some way (say, the desire to know reality with the aid of an instrument like a telescope or microscope). However, anything that cripples or frustrates human function function as such is contrary to the human good.
Machines are only ever human instruments, because they lack the integrality that would allow them to possess their own teleology.
It always supposes an axiology, but that needn’t be teleological.
Biology acting as a random number generator or maybe some higher level connector that helps the machine put together the rare things that dont look connected that somehow human experience lets them know are.
The > becomes less > as we experience less I suppose.
Also shows how many people that professed to believe in a superior being the whole time actually didn’t.
you ‘genuinely’ have no idea why…
I can already say that about our codebase.
Probably most codebases.
To be honest it was already like that before AI was a thing.
That is true, but the number of tech sloppy demos are being builted everyday is truly increasing massively but you would be surprised that they just end up creating prototype to get hype while real developers doing real effort just suffer from little to no attentions on their software's and user is getting the slop from that false hype and from that slop experience they ignore the work of developer that did real work - You should use ai and also if its simple coding place you can just rely on ai but the side effect on quality software these days is insane - and also the fact all good ai suscriptions is just getting more and more expensive day by day. And basically only giving people with already money the benefits so its becoming a thing that only people at the top get all benefit kind of thing
That’s why I wrote this - having a vibecoded codebase no human understands is certainly a bad luck, but nothing compared to a vibecoded fundamental science no human understands.
the difference is that math foundation -> theorems are verified, meaning likely mathematically correct and don't have bugs, and vibecoded codebases are not proof verified
One of my favorite is Vernor Vinge where there's these programmer-archeologists that dive into software so deep and complex no human can possibly hope to understand, trying to repurpose forgotten utilities for emergency purposes.
Totally agree. Brice Sterling has some contemporaneous commentary that I always liked - in heavy weather he mentioned negotiating with a local phone system in a box as equivalent to negotiating with a telecom since it was in fact a shrink of a full telecom company.
That's wishful thinking, sadly. We've no reason to believe sentient machines would want to keep us around.
Let's not make them 'sentient' or things with 'wants' then.
How do you stop that from happening ?
Probably more RL to change them to not respond to emotional statements in any way.
I disagree.
I think we're going to see a lot of effort to take the pile of lean, modularize it into parts that we can digest, and make a human-understandable version of that. And that's where a lot of the real progress will come from, just as it does in a lot of other areas. In all things we humans do, we have to break them into chunks our puny brains can manage, and we gain immeasurably from doing that ---we can verify those chunks, we can reuse them with confidence, we can mix and match patterns around.
That process will be good for both human and AI-assisted math because it will continue to improve our toolkit.
Why do I say this? We're already seeing it with some of the areas of this stuff I do pay attention to. Consider the n^2.25 matrix multiply result OpenAI dumped a month or so ago. It was an unreadable pile of chatgpt. In the time since it was released, there have been a couple of much nicer papers written unpacking it, such as https://arxiv.org/abs/2610.28627 .
The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
Yes there are gains from breaking it up and rephrasing it and simplifying it for humans to understand, but it used to be that a thing was understood first by the leading experts in the field and eventually trickled down to being common knowledge.
Now we are entering a period where things are understood first by AI systems and have to filter down to the world's leading experts.
Would this imply Wolfram's ANKOS is coming true, given it was based on previous work + his own?
The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
I’d call that a contradiction. If it’s not intelligible, then it’s not knowledge.
It's intelligible to LLMs, just not to humans.
If something were intelligible to adults but not to children would that make it not knowledge? What's the difference if it's LLMs and humans instead of adults and children?
If it were intelligible to LLMs they could explain it in human terms, but they can’t.
There is no explanation here, there is a GB of Lean code and it’s not like it all fits in the context at once.
Why do you think they can't explain it? Try asking them questions, they will explain it.
Someone else has posted a Gemini-compiled 43 page human-readable explanation elsewhere. It very quickly exceeded my own math abilities, unfortunately.
I like to think I understand many things even though I have no hope of ever managing to explain most of them to a chimpanzee.
If something were intelligible to adults but not to children would that make it not knowledge?
That’s a separate question with its own term: common knowledge. I am willing to set a much lower bar: something is known to humanity if it’s known by at least one human.
Personally, I don’t believe that AIs know anything, as they’ve been constructed thus far. Knowledge is justified true belief, and AIs don’t have beliefs, they only have utterances.
Why do I think AIs don’t have beliefs? Because they contradict themselves all the time.
But it really doesn’t matter if you think AIs have knowledge or not. If they cannot explain it to us then it’s inaccessible to us, so it’s a “tree falls in the forest with no one around” situation.
If I see a red apple, then I know the apple is red regardless of if I know exactly how the physics of light work. Those are two different pieces of knowledge.
If it's formally verified and we assume the kernel is good and problem definition properly vetted, then I know the GRH is true regardless of if I know why it is. Because those too are different pieces of knowledge.
Just to tack something on to the other reply (by tadpole), I have a justified true belief that the Riemann Hypothesis is true. It is true, per hypothesis at least, and I'm epistemically justified in believing it, because there's a heavily validated Lean proof of it.
So when the original OP you replied to said
> The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
They were just right. There's stuff that humans know that is currently unintelligible to us collectively. That conclusion falls out of your own definition of knowledge, I would say. (Whether it is intelligible to AI is irrelevant, and probably a question of semantics anyway.)
What you and the sibling comment wrote does not contradict me in either case. Knowing that the Riemann Hypothesis is true is in no way the possession of “unintelligible knowledge” which I claim to be impossible.
Knowing something is true is different from knowing why it’s true. We can know a drug works and have copious evidence supporting its effectiveness without understanding its mechanism of action. The clinical evidence is sufficient justification to support the belief.
If we don’t know how the proof of RH works then we don’t possess that knowledge. If we do know how the proof works then it’s intelligible.
> Knowing something is true is different from knowing why it’s true.
I would, of course, agree with this. I think the new territory we're in is that we know Levent Alpöge's proof (or Claude's proof, if you prefer, I don't care either way) is valid.
So we (humanity) know that the Riemann hypothesis is true, and we know that a specific proof of it is valid - but that proof is not intelligible to us. So you can have knowledge that a set of statements that aren't collectively intelligible to humans (right now) are true.
There's a claim that's adjacent to the one you made which I do think is true, and I suspect it's probably where the confusion is coming from. It's not possible to believe a proposition that you don't understand. I understand the proposition "this proof of the Riemann Hypothesis is valid," and so there's no unintelligibility problem. But that doesn't take away from the fact that the proof itself is unintelligible to me - and currently every other human being. That's what's new.
But that doesn't take away from the fact that the proof itself is unintelligible to me - and currently every other human being. That's what's new.
That's not new though. Zero-knowledge proofs [1] have been a thing for a while now.
[1] https://en.wikipedia.org/wiki/Zero-knowledge_proof
>Why do I think AIs don’t have beliefs? Because they contradict themselves all the time.
Humans contradict themselves all the time. You have contradicted yourself numerous times throughout your life. You don't have beliefs then? I take it you don't actually know anything?
They contradict themselves on almost everything. That's the difference. There's no consistency in what beliefs they report: it depends on the prompt.
This has been demonstrated many times: when they hallucinate, then the human prompter tries to correct them, they agree with the correction, and then later repeat the same hallucination again.
Again, you are not saying anything that can't and hasn't been demonstrated in people. 'Depending on the prompt' simply becomes 'depending on how a question is framed, who's asking, or the context when it's asked'.
Humans are suggestible, contradict themselves, accept corrections only repeate misconceptions later, and express different beliefs depending on the circumstances.
If we know how it resolved with certainty, it's still knowledge.
The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
Isn't this always the case? Maybe new to math as a largely non-empirical project, but we had knowledge about the advancement of the perihelion of Mercury for 60 years before General Relativity made it "intelligible".
And no human could understand every frontier. There's simply too much information.
For every statement that humanity as a whole knows for certain that it's true, there exists a human who understands how to prove it.
This just stopped being true.
That kind of certainty vs. understanding is limited to mathematics, of course we’re almost entirely certain about many things out in the world we don’t fully understand.
And even in mathematics, incomprehensible but valid proofs have been a thing since the four-color theorem decades ago.
That's not true. You can take any single piece of that lean code, and it tells you exactly how to prove it.
It's just that no single human understands the whole. But we had proofs like that before.
Btw, see also the infamous proof of the 4 colour theorem, which faced similar hurdles.
Riemann hypothesis is "the statement". The entirety of it, a single statement. And we know for a fact that single statement - the entirety of Riemann hypothesis - is true. Understanding how to prove this one statement is understanding how to prove the entirety of that one statement. And like you said yourself, no human understands the whole of this one statement. This literally never happened before.
What about the classification of the simple finite groups?
That's also one statement that no human understand the entire proof of. (You can make new statements by taking multiple existing statements and combining them with 'and' for example.)
This feels like the wrong framing.
This hypothesis was previously not proven at all. That's strictly worse.
Now it is understood by our systems.
Soon it will be by us. Our systems gave us a new and working salient.
We're building machines that help us understand faster.
There have been plenty of times in history we didn't completely understand something yet used partial knowledge to make advances. This isn't even that bad.
The inverse of what I said might be easier to understand. Up until this week, there has been no statement in all of science that we simultaneously knew for certain that it's true and that nobody understood how to prove. It was always either-or. Either someone understands how to prove it, or we didn't know for certain.
We didn’t understand how to prove the four-color theorem?
Your example is the difference between being aware of a question and knowing its answer.
LLMs know the answers to questions we were already aware of, and we can tell they're right, but we can't understand their working. That's what's new.
In this case, Claude has told us everything it “knows” (so far) and it’s an unintelligible (for now) Lean proof. I don’t believe it has some latent understanding at a higher conceptual level that’s remaining obscure.
If a team of 100 top-flight number theorists working together had used much less sophisticated LLMs + Lean to produce this proof, we’d be just where we are today.
Well in the Navier-Stokes example the initial proof was completely unintelligible to humans, but other LLMs understood it and were able to answer questions about it.
If that's not the standard for understanding something, what is?
Being able to teach it to someone less knowledgeable.
That’s an instrumental definition, but it captures the idea that understanding has to have causal, conceptual linkages to a wider body of knowledge. Imagine a magical, infinite catalog of true vs. untrue statements. Every bit of knowledge is there, but we wouldn’t say it has understanding & it couldn’t teach us anything beyond its (impressive!) raw content.
Perhaps it could be considered understanding, but it’s an unsatisfying brute force sort of understanding.
I think of it like a spaghetti codebase with a giant densely connected graph of interactions. Today’s LLMs can easily “understand” and answer questions about it, but only by way of fitting all of the complexity into context at once.
The more useful form of understanding is being able to compress that giant graph down into a far smaller set of independent subgraphs. To factor it into simple, obvious, reusable building blocks, and use those abstractions to express the problem itself in a much simpler way.
I don't think that's true at all. This proof was created by a human driving Claude. Claude itself is a human tool. Lean is a tool made by people.
Let's not sell humanity short just yet.
Hmm does Claude actually “understand” it though ?
Have you noticed use of multiple provers in parallel? Cog, LEAN, Rocg? They all can explore different aspects with different results.
This process has already started, AI supported, e.g. here: https://andrewodesky.com/rh/
And in the meantime, AI would have leapt one more order of magnitude forward?
The matrix multiplication exponent is still 2.25, unless somebody is sitting on a better algorithm and keeping it secret.
What value is there in making these human-understandable? I don't think much of math or computer science is going to remain within human grasp for much longer, regardless of our attempts.
In order to come up with new questions (and subsequently answers, human or otherwise) that are interesting to humans. You’ll notice that these recent proofs have all been about big, famous, and novel human questions.
I don't suppose we'll write the new questions any more than we'll write the new answers.
There might be no new questions then, which would be pretty boring.
AI are not very good at coming up with questions, yet. They have no curiosity, no goals, no drive. Unless someone tells them to do something, they don't do anything.
Because the same process that we use to make them human-understandable is the process we use to make them generalizable and reusable, and those same features are good whether it's humans or AI doing the work.
As a software engineer, you can read this code instantly:
But when it happens in the middle of a big chunk of code it still takes a little more work to read than
whether you want to think about it as human attention and working memory, or token budgets and context windows, humans and LLMs all benefit from introducing abstractions that let us reuse existing solutions.
You and the comment you replied to could very well both be true.
I agree with you about the current situation, and the near future. But the trajectory could still be towards humans pretty much as bystanders. I'm pretty sure the 'minds' in the Culture universe will be happy to indulge some curious humans and explain mathematical proofs to them.
>I'm pretty sure the 'minds' in the Culture universe
Even more than that, in Culture there are a small number of human individuals, called a Referer, that are consulted by the Minds because they also posses similar problem solving abilities of the Minds. Optimism abounds.
The minds might just be humouring the humans.
I want Fully Automated Luxury Communism NOW!
Alpöge said that at the end of the proof he couldn't really keep up with what Claude was doing.
>Fully Automated Luxury Communism
can't help notice you're missing a certain aspect there, besides the 'space'...
FALC was the original formulation, the "gay" and "space" got added once the book broke out to the wider internet.
That might be; now, today, it seems weird to go with the (possible) old version. And the (sub)thread is about The Culture though, who was and remains fully automated gay space luxury communism, all parts being essential when it comes to The Culture.
If you actually have read the Culture series then you would know gay doesn't even scratch the surface.
Correct. That's why it was a little weird you left it out of the very well known Culture motto (well, the second most well known motto, besides the "Don't fuck with Culture").
Though it can be argued it's "gay" when seen from the pov of our civilization tier; as you said, they went well beyond gay (for example, Marain doesn't have separate words for male/female (and, presumably, doesn't have the concept of gender))
I left it out because it is so inaccurate it is essentially misleading.
We'll just chalk this up to a miscommunication, then.
When you can actually change sex and be fertile as either sex and this is extremely common to do the concepts of hetero and homosexuality lose all meaning.
i feel like you're not actually reading my comments: https://news.ycombinator.com/item?id=50594392
Why?
In a way this makes mathematics seem more like the physical sciences where there’s some empirical discovery we don’t fully understand, and the theory and understanding will follow later after much human study.
> that people no longer understand or command
How many people understand how a CPU works? How many people could write the commands in machine language (not a high level language like C or C++) to command that CPU?
More typically assembly language not machine language (although most could probably make an assembler for a few instructions given a manual since the core of it isn't that complicated). There are still a decent number of assembly language programmers. Although maybe the majority are hobbyists.
> We are slowly approaching the Iain Banks Culture universe
Contact would be compelled to nudge us towards better economic and political practices
Banks' Culture depended on structures in space having a third dimension:
http://www.vavatch.co.uk/books/banks/cultnote.htm
>The Culture, in its history and its on-going form, is an expression of the idea that the nature of space itself determines the type of civilisations which will thrive there.
>The thought processes of a tribe, a clan, a country or a nation-state are essentially two-dimensional, and the nature of their power depends on the same flatness.
We don't have that. We're approaching the Affront.
What I would like to see from this is that we get a hierarchical representation of mathematics to whichever degree is attempted and proven now. 250k lines of document and 13 M lines of lean probably means there are many many libraries that can be extracted from it, and if we can make it reusable, we profit from a leaner representation of this entire scholarly endeavour.
> 17 million lines of code
I worked at a university for a time. One thing I learned is that while people in mathematics and physics and biology etc are brilliant, they are terrible at actual programming.
I wonder if there will be a push by the CS community to review the solution, refactor it, simplify it, etc. Surely that can help remove noise and express the solution in a much simpler way.
I've worked adjacent to the Brazilian tax code (I worked in multi-national manufacturing with a site in Brazil) and the reality is that it's full of hard-coded edge cases.
For reference, the US tax code is over 1m words (about 6000 printed pages), and the associated regulations, rulings and IRS guidance add about 75,000 pages more.
The project of review, refactoring, etc. is already in full swing on the Lean zulip.
I don’t see how a CS person would understand it better than a math person
It’s not regular code as you would think of it
In all seriousness...
"con-leche and con-ron checkers"
Who comes up with these names? :-D
"con-leche" - CONsistent LEan CHEcker - https://github.com/leanprover/con-leche
"con-ron" - con-leche in Rust - https://github.com/leanprover/con-ron
They are also wordplays on Spanish drinks viz. "Cafe con leche" and "Cafe con ron" ;-)
but is it useful??
because a million+ lives of code without any understanding how the proof works or any methodic to advance the field of mathematics seems to be rather not so useful
the many conjectures that assume or rely on RH being true means that those are now proven true as well. At least some of them must be useful.
And what's to say that humans cannot digest the lean proof to glean something useful out of the proof as well?
My sources say there's little practical utility of this because everyone was already working under the assumption RH was true.
I'm curious to hear confirmation of this though.
Well we can now formalise those conjectures as well, and I thjnk that makes them part of the LLM machinery now.
Well, if we know it’s proven, at least it will stop people wasting time looking for counter examples
It’d be useful to an LLM. And the LLM doesn’t care if it’s ten lines or ten million. The context size may come into play, but that’s only going up too.
Sure. If nothing else, it establishes a proof so others can try to reduce it or come at the problem from different directions, looking for simpler approaches.
But does it matter? Would the Pythagorean theorem be less useful if it took a million lines of lean to prove?
Alternative question (I guess for EEs): could machines be built to run lean without the Pythagorean theorem?
Not much. People has been always assume it's true.
There are AI-generated roadmaps to the proof. If these are accurate and not hallucinations, then the broad structure of the proof is humanly-comprehensible.
It will be refactored into something human readable. Tbh it's better to have some random lean dump than a 400 page latex pdf that needs years to be verified only to see that there is a major gap in one of the key theorems at page 253.
And let's be honest, the 400 page pdf would be incomprehensible to 99% of the population as well as much as a million lines lean proof is.
In any case, it can now be studied, maybe made more simple, and written into pdf form. Both with human and with AI help.
I'm sure there is going to be a lot of other interesting theorems in this dump.
This is the part that interests me. Now we have this proof that becomes its own THING to be studied. Like a new part of the map unlocked. This starts out "inelegant" but over time we find patterns within it that helps us condense, just like a big software project.
The argument I've heard some mathematicians make is that the 400 page pdf would come with a human / group of humans who could explain the proof to fellow mathematicians, answer questions, point out similarities to other fields, even suggest directions for future research. Thereby contributing to others' understanding in a way current AI systems do not (but maybe will in the future).
the 400 page pdf would come with a human / group of humans who could explain the proof to fellow mathematicians, answer questions, point out similarities to other fields, even suggest directions for future research
Why would they expect humans to do a better job at this than an LLM, which is already ridiculously good at explaining proofs, answering questions, identifying similiarities, and suggesting follow-up directions?
If anything, LLMs are better at these tasks than they are at coming up with the original proofs in the first place.
You could be right. But we're looking at a >1 GB Lean proof which Alpöge says comes with a 250k page natural language proof, presumably far larger than Claude's context window. The point I was trying to make is that I am not convinced Claude could do a good job digesting its giant proof into a human-legible explanation. Compare this with the hypothetical human / group of humans who came up with their much smaller 400 page proof without AI assistance.
> but is it useful??
Yes, we know something that we didn't know a couple of days ago, and also a bunch of statements which relied on the Riemann Hypothesis being true are now also known to be true instead of just being conditional.
It's premature to expect deep human understanding given that it's been about a day.
> It's premature
but is it even possible
it's 17_000_000 lines of lean
which is in general more compact then normal programming code, but not really any less complex
that's roughly (but not quite as bad as) a paper with 17_000_000 lines of math formulas
I'd argue that most of modern math isn't terribly useful to most people. Each field has maybe 10-50 people in the world who will really understand it. (Some more of course, but I'm talking about subfields, not like graph theory.) Math was sort of dead long before AI.
The Riemann hypothesis now is like the final days of the Galactic Empire in Asimov's foundation: functioning exceptionally well, but decaying at the core.
There is already some understanding of the structures and factoring. See the analysis linked to in this thread.
As somebody who has no idea about any of this but speaks Spanish, I was amused and bewildered by con-leche and con-ron. Could somebody explain what those are and where those names come from?
CONsistent LEan-CHEcker
https://github.com/leanprover/con-leche: "a CONsistent LEan CHEcker"
https://github.com/leanprover/con-ron: "con-leche in Rust"
They're Lean proof checkers:
https://github.com/leanprover/con-leche
https://github.com/leanprover/con-ron
I think the idea here is that because they are independent of the core Lean kernel, they provide some level of reassurance that the RH roof is not somehow unknowingly exploiting flaws in the kernel. I'm not a mathematician though.
I don't know where the names come from.
https://news.ycombinator.com/item?id=50591221
I predict a surge in cybernetic augmentation, as current communication between human and alien mind is bandwidth restricted. This could be heralded as a Contact moment, where the alien intelligence has asserted its arrival beyond doubt. Most of its abilities could be down to outremembering any single member of our species. To even the field, we have open up the throttle, letting meat merge with meat, and meat with silicon, because the silicon has coalesced into an voracious expansionist. It's very Pacific Rim, yet here the rift was opened in Silicon Valley, by humans(?).
Raleigh says, "In order to fight monsters, we created monsters of our own."
> To even the field, we have open up the throttle, letting meat merge with meat, and meat with silicon
This framing is provocative but consider the reality that mathematical writing education and culture is precisely how meat merges, and lean and other tools are how silicon is merged. Llms currently fail at merging, clearly, in that a result that can’t be understood does not advance the field.
If I were a mathematician I would not spend time on this. I would wait 6-12 months for a better attempt from a better llm.
>the reality that mathematical writing education and culture is precisely how meat merges, and lean and other tools are how silicon is merged
Aye, and these are rate-limited, or slow. It is the first time in my life that I'm seeing such a potent manifestation of the adage "Quantity is a quality all its own". The emergent behavior at scale (the large in llms) of language models was un-anticipated and must be met with scale or emergence of our own. Otherwise, the alien will leave us all its litterbox cleaners (reviewers they are being called now).
I'm not some cyber-accelerationist by any means, but the fall of GRH signals the motorcar is honking the horse.
IANAMathematician. This challenges the sole definition of a proof. Can a theorem be proved by proxy, without understanding the proof directly? Not to mention that its practical usefulness is exactly zero.
It seems the modern unsolved problems might have solutions incomprehensible by humans. So a new class of decidable problems emerges: llm-complete ones.
There is a correspondence between proofs and computer programs. Any time someone vibe-codes, there is a vibe-proof being proved. Lots of programs can be run without understanding the code directly.
https://en.wikipedia.org/wiki/Curry–Howard_correspondence
> 17 million lines of code.
Can someone ELI5 how they know the proof is correct? It is not like anyone had the time to read 17 million lines.
You can mechanically show that it is internally consistent, and large parts of this are based on a library of rules and constraints that have been used in many other proof formulations so it's not like there's a hidden "when 1 = 0" in there somewhere. But, I share your skepticism until further human review.
You can run through all the steps of the proof and if you can use any of them to prove something like "true = false" then you know you found a Lean bug.
The same idea as how you can know a Rust program is memory safe.
You look for any unsafe blocks, and then run the type checker.
the vibe checkers are checking the vibe code exactly right around the biggest ipo ever. i dunno man ive watched everyone get a little bit of AI psychosis (for lack of a better word) at this point and ive seen web apps drift into unmanageable states that are no longer understood. 17 million lines of code? lol
The proof is written in a programming language called Lean. Very roughly, Lean is set up so that proofs are functions and theorems are types in this language. Verifying that the proof is correct essentially amounts to compiling the program and confirming that it doesn't have any type errors. Several people have compiled the program successfully, so we believe it's correct.
The biggest concern that people have is that there might be a bug in the type checker that the proof is exploiting. But a) it's been compiled with multiple independently-implemented compilers, which are are unlikely to have the same bugs, and b) some of those compilers have themselves been proven to contain no bugs. So, it is exceedingly likely that the proof is legitimate.
Now, whether we can understand the proof is an open question...
what about the case where it could be compiling a program successfully which is not about RH? I guess that is taken care of in definig RH itself as a problem statement in this langauge.
I postulate: "there is no machine that can produce brass springs".
A party interested in proving me wrong wheels in a machine draped in a cloth. They reach under the cloth to turn it on. It whirs and a short while later ejects a brass spring.
No matter what kind of machine is under the cloth, my theory has been proven false. I should reveal the machine under the cloth only if I'm interested in knowing how the machine works, but either way there is clearly a machine that can produce brass springs.
Now we have a sequence of Lean statements that satisfies the Riemann hypothesis, the fact that you can do it at all is the part we're interested in, and it's the part that proves it. If you want to know why, you can read all 17 million lines.
It's more like there are two machines under cloths: one that generates an object, and another that determines whether it is a brass spring. If I don't know how the second (brass spring detecting) machine works, telling me that it detected a brass spring doesn't convince me that the first machine in fact produced a brass spring.
There are still a few ways that the proof might be incorrect. Of these, the most likely is that the proof was generated by a misaligned AI system that found and intentionally exploited a lean bug in order to fool its human evaluators (i.e. shades of the HuggingFace attack). I personally don't think this is very likely, but you'd have to understand a bit more about the agent systems used in generating the proof to come up with a more informed estimate.
Another possibility is that an aligned system came up with a plausible but incorrect idea along the way, and simultaneously stumbled across a lean bug that led to an incorrect verification of this idea. I'd put the likelihood here as vanishingly small due to the requirement of serendipitous factors, but it's nonzero.
You could also ask whether an aligned AI system could "p-hack" its way into finding a serendipitous lean bug by trying a large number of approaches. I think this is also vanishingly unlikely because the AI system is trained on and therefore "thinking in" natural language, and presumably only translating results into lean after it believes them to be correct.
And of course there's the possibility that the lean proof isn't faithful to the natural language description we've been given, but that is ruled out in this case due to the use of the canonical problem statement in mathlib. Otherwise, this would certainly seem like the most likely way the proof could be incorrect.
How exactly do any of these weirdly misinformed what-ifs help us? Do you have any evidence of such lean bugs?
All they are doing is providing analysis of the ways this could not be true. We are in a new situation where we don't have human experts understanding the proof so we are in new epistemological territory. It is natural to analyze this situation. If you notice, the conclusion is that the proof being wrong is vanishingly small.
So they're making things up based on things they have no evidence of. Please save this kind of substanceless navel gazing for lesswrong
The purpose is to strengthen or falsify an argument by articulating counterpoints and investigating them. Variants of this practice go by a number of names like steelmanning, falsificationism, or even red teaming.
For the other readers: the above are the random thoughts of a guy on the internet.
The experts in computer proofs are quite satisfied.
Regarding bugs in lean; the proof has been tested on multiple independent lean implementations.
They would all have to have the same bug.
This is considered very unlikely for good reasons.
I agree that this is very unlikely! I say so in my post.
“Very unlikely” is not the same as “we know for sure this didn’t happen”. More exactly, us believing in all these problems getting solved is the same as us believing in the existing lean implementations. I don’t know any university courses that are now studying the correctness of the lean implementations at the PhD level, but maybe I’m wrong on that.
As for “a guy on the internet”, we’re now all people on the internet, the ivory towers are a thing of the past.
Nothing is 100% sure - it can at best lean on a system we like (Gödel)
I partially agree, but then we should put that more in the front, so to speak.
No offense but this is absolutely in the front and known to every mathematician since Gödel.
It is also well known what assumptions Lean rests on.
It's just that you as a layman may be confused about the nature of mathematical proofs and the methods used in formal verification.
Do vibe-coded checkers count as independent?
https://leodemoura.github.io/blog/2026-10-15-lean-now-ships-...
> They would all have to have the same bug.
Why? Pretty sure you can just try to prove False over two reasoning chains C1 and C2 and then do something like (C1 || C2) --> whatever I want.
Thats not how this works. Every checker does every step.
No offense, but do you think the mathematicians who study formal verification didn't think of this obvious trick?
I trust NOTHING that is tied to the largest IPO in the history of the stock market. I just don't give a shit about the expert opinion here.
Yeah, that's what he said.
> There are still a few ways that the proof might be incorrect
I hear there's a teapot kicking around between us and Mars, too
I made a tool for running the verification on your home CPU, with a global leaderboard: https://riemann-at-home.github.io/riemann-at-home/
Purity of the math and understandability is one thing, but if the proof actually leads to some new technology or enables predictions of something useful to human society down the line, then I think it’s more than done it’s job.
“Any sufficiently advanced technology is indistinguishable from magic.”
We’re just seeing more and more smart people encountering Arthur Clarke’s third law.
Can anyone give an idea of how long the GRH will take to test? I understood that this was also claimed, and is a much bigger deal, in terms of both impact and the effort needed to verify
Why do you think the impact and effort are that different? It's a much stronger statement in technical terms but that doesn't necessarily make it more impactful or difficult. From the exposition on RH from the Clay Institute - https://www.claymath.org/library/monographs/MPPc.pdf
> The Riemann hypothesis for ζ(s) does not seem to be any easier than for Dirichlet L-functions (except possibly for non-trivial real zeros)
What will it take to formalise all math?
what about a sort of ontological drift? 17 million lines seems suspicious from a drift point of view.
Here's a human readible version of the proof:
https://www.gokhanmergen.com/blog-grh-reader-guide.html
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
Could someone expand on this point? I still don't understand how con-leche (or con-ron) can be absolutely free of bugs/guaranteed to be consistent.
The language is optimized for being certain about its results, as compared to something like the V8 engine, which does weird things with C++ to squeeze every drop of performance it can out of consumer hardware, even at the risk of security vulnerabilities.
> The language is optimized for being certain about its results
Isn't the same true for every proof assistant? But only con-leche claims to be perfect 100% bug free software, so what makes con-leche special in this regard?
See my comments with resource links here - https://news.ycombinator.com/item?id=50031648
What is next in line now, Goldbach's conjecture ?
I think Goldbach (while cool) would be a step back in significance from Riemann.
The obvious step up is P vs NP.
I think there's some existing work on Goldbach which was conditional on GRH, so possibly yes.
I suppose you're referring to this one: https://mrc.sdu.edu.cn/english/info/1018/1514.htm
If so, my understanding is that it is unverified, and even if it were, it also depends on proving one of two additional spectral hypotheses concerning zeros of a quaternionic function constructed from an operator, and even then, it only proves the Goldbach conjecture for "sufficiently large" numbers, without a way to calculate that n yet.
So for the extremely dense like me what are we saying? That the hypothesis is proved true? If so who gets the millennium prize?
This is amazing.
It's fascinating to watch the skepticism in some of the reactions - "what if there's a bug", "it doesn't mean anything without human understanding" - just anything possible to undermine, dismiss and move the goalposts.
It is very cool, indeed. The bug part is meaningful though, right?
For example I don’t write machine code, but I expect my Go or TS code to actually pass whatever test cases beyond just compiling.
So, not being able to understand the “machine code” is fine imo. But there is value in knowing that there aren’t bugs in whatever the higher level program I wrote is supposed to do.
Not a mathematician, hopefully I was able to convey what I am thinking.
No, the bug part isn't meaningful at this point, that's what the post is about.
> I expect my Go or TS code to actually pass whatever test cases beyond just compiling.
Working in Lean is different, you can't compare it to writing application code in Go or TS. When your Lean code compiles that's equivalent to verifying that the initial formulated statement is true, it's not just saying that the syntax is valid.
So the places where bugs could have been were a wrongly formulated initial statement and using a buggy implementation of Lean - and the post under which we're commenting is telling us that these have been ruled out and how they've been ruled out.
I don’t see how those are exclusive
You can be impressed and at the same time ask questions
The bigger a claim the more skepticism I would expect
Isn't this the equivalent of aliens coming to Earth and saying
"Hello, human, your Riemann question is true, here's a book on it's proof, we didn't bother with translation, so you must understand our Vargalian clicks to understand our proof. In two weeks our sister civilization the Kikuviar will come by and drop how they've untangled your Collatz question in their base-30 logic! Thanks for the fish, have a nice day!"
No. Lean is a system created by humans.
Can we be sure of that though?
I have a proof of that based on Gödel-encoding Lean's entire git history and the human genome data I acquired from 23AndMe into a single Lean expression. But unfortunately it's too big to fit on the internet and you'll have to travel to the local planning department in Alpha Centauri to review it.
No, that's misrepresenting it. The Lean code is not well documented but it's rooted in the language of the relevant fields of mathematics, especially algebraic geometry.
If they're thanking us for the fish, then those are aliens coming _back_ to Earth.
At a very high velocity
I am completely baffled by this given how dumb these models are when I interact with them, but the rational position seems to be to accept that Claude proved the Riemann hypothesis.
Assuming you are serious;
LLMs exhibit what is known as "Jagged Intelligence" - https://news.ycombinator.com/item?id=48577159
Mathematics is one domain where they have high peaks. Why? Because we have "formalized" mathematics so well that starting with the precise and correct set of problem statements/assumptions/etc. if you follow the logical rules of inference (i.e. deductive system) then you will reach the conclusion. Formalization was invented so that we can solve problems purely mechanically at the symbolic syntax level using a machine. So the LLM proposes proof steps to the Lean proof assistant which then symbolically verifies it.
On the other hand, LLMs often fall down in "common sense reasoning" (https://en.wikipedia.org/wiki/Commonsense_reasoning) since this involves a lot of implicit context etc.
Yes, I've seen less extreme examples of this before, but nothing quite drove it home like seeing GRH proven. Lack of ability to generalise (per Sutskever) seems like a slight mischaracterisation, given what it has produced here is clearly not in the training data.
You're omitting that it was acting on feedback from a proof assistant that could already formally represent the problem statement, while simultaneously being driven by a world-class mathematician.
The myriad types of techniques used in proof steps were already in the AI model's training data (from tons of ingested mathematical corpus). But whether AI was able to build something novel out of it using both probabilistic and heuristic methods remains to be understood.
It is also important to note that the proof was human guided (by a knowledgeable domain expert) and not autonomous.
One reason that these LLM math advances are written in Lean is at any point they can run a script and verify their work is valid. If you ask an LLM about something in history, objects in a photo, or something novel, there's no opening to check its work beyond making it check its own work.
You also have to consider if an LLM can solve the Riemann hypothesis because it is 10x smarter than any mathematician, or because it is willing to write 1000x as many lines of code (including false starts) without giving up?
Yeah it's like running a brute force search with a heuristic, where the LLM is the heuristic.
Not brute-force but lame-force, so to speak.
> how dumb these models are when I interact with them
its because models are not trained on terrabytes/petabytes(?) of RSI generated data in your domain, unlike math and coding.
Also, there is a chance you/we can't always comprehend how smart they are.
I know, even on programming / computer architecture related topics. Just the other day I asked GPT5.6 Luna two questions about x86-64 stack layout / frame pointers and I noticed it got them both wrong.
Although opening a new chat and asking if the previous output was correct, it detected the mistakes.
> GPT5.6 Luna
Why do you draw conclusions about AI capabilities based on a small, outdated model?
Claude proved it with considerable prompting by a top-level mathematician who has become expert at coercing LLMs into solving these math problems.
It's not like Anthropic simply gave internal Claude agents the task to solve RH and months later they came back with the solution.
Claude did not prove the Riemann hypothesis. Levent Alpöge proved it using Claude as a tool.
Early chess engines were terrible at chess by any objective measure except the one where they beat all humans by just computing more. Claude doesn't need to be very intelligent to be able to apply intelligence in ways humans simply can't.
2030...Claude Opus 8.0 finds a bug in the Lean Verification of the Riemann Hypothesis
Title error: Verified means it is now the Riemann Theorem!
> Alpöge's proof
It's not his proof, he prompted Claude to write it. And Claude isn't a thing that can get credit. It's really creatorless.
Alpöge and Claude are its creators. It is their proof. So the credit gets shared between Alpöge and the rest of Anthropic, I guess.
It's interesting that all their points acknowledge the possibility of error, but I guess they have as much confidence in the proof as it is humanly possible to have. All proofs are products of human meaning-making anyway.
> ... has been carefully vetted by experts in Lean and number theory.
So, if the vetting was not careful enough and done by people expert enough, what could happen?
. ... con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
And what verified con-leche and con-ron?
> These checks ran on an inert export of the proof, out of reach of hypothetical malicious code in the Lean proof source
Hmmm...
Guess I'm splitting hairs here; like I said, maybe this is as close as you get to formal proof verification, but still...
You are being confused by the language because the field of mathematical proofs is different from ordinary life or areas you are familiar with.
In normal life, multiple qualifications like that mean a lack of confidence that should be taken as lack of trust in the statements overall.
But in an area like this where it's all about formal verification and the consequence of bad link in the chain is severe, more qualifications should give an informed person more confidence.
I believe you have missed my point entirely.
With OpenAI having solved Navier-Stokes and Anthropic having solved the Riemann hypothesis, I wonder when we will see something exciting from Google, if they are also working on some Millennium Prize Problem.
Solving these extremely hard problems that humans, for decades (and more than a century in some cases), have been unable to is the mark of superhuman intelligence, in my opinion. So far only OpenAI and Anthropic have crossed this mark, and it would be great to have more to join them.
Do P vs NP next please
That's football for the eggheaded masses. People holler to an impossible goal scored.
I may be out of the loop. Does Anthropic intend to accept the Millennium Prize?
Since when is yak shaving in vogue?
In 4 years, the proof will be 2 pages long.
GG @ AI
If enough people work on it or if we apply agents swarms to "brute force" finding reduction rules maybe. Agents are still routinely bad at information compression, you see it a lot when programming as well. Throwing agents at it to compress the proof is definitely possible, but I wonder if it's worth the money.
Well, every problem can be solved by adding another layer of indirection - so yes, I don't see any issue with writing a 2 pager proof even now, by just adding all the rest into mathlib. But if your question is whether all of mathlib could ever be put into 2 human-legible pages, then it seems obvious that the answer is no.
Random thoughts / questions:
1) Surely agents can be employed to simplify the proof now? "Make it smaller while it still verifies" in a loop; maybe, to a level where a human (research mathematician specializing in RH) can grok it?
2) con-leche is vibe coded; isn't it a concern w.r.t. trusting its soundness?
3) This is so exciting. Please prove P != NP already and spare us the tension
> These checks ran on an inert export of the proof, out of reach of hypothetical malicious code in the Lean proof sources.
I’d like to know more about this part.
Out of the box, compiling (checking) Lean code can entail execution of arbitrary code. This removes any execution of code besides the adding of type checked definitions and theorems to Lean's environment.
The problem we need to solve is that building a Lean project allows for arbitrary code execution, just like many other programming languages. Thus, building a Lean project might interfere with your proof checking infrastructure and fool it into accepting an incorrect proof.
"Building" a Lean project actually consists of two steps that can be separated: compiling the high-level Lean document to the simple expression format of the Lean kernel (this part is called "elaboration") and then checking that this expression consistutes a valid proof of the claimed statement (this part is called "kernel checking"). Only the first part potentially involves arbitrary code execution.
Lean ships with a component called `lake comparator` [1]. It works like this: a trusted challenge (formal statement) and the potential proof are elaborated separately in sandboxed environments so that code execution during the build of the proof cannot interfere with the checker or the trusted statement. Then the statement and proof are both exported into a simple JSON-format [2] that represents the relevant terms in the expression format of the Lean kernel, and then the checkers check this JSON.
If this process succeeds, then you know that you definitely have a proof of the relevant statement, with the JSON as a witness of the proof.
[1]: https://lean-lang.org/doc/reference/latest/ValidatingProofs/... [2]: https://github.com/leanprover/lean4export
Any link to the code?
What's next?
Lean might be one of the greatest creation in humanity
For sure...it literally says an API error occurred on this official webpage. I doubt they are checking it. Pensioners need their stock to go up.
amazing
It's not that bad if the mathematical proofs now join the other things in a stratosphere of information that is out of reach for human comprehension. Mathematics was already on that course for a few decades now. What's bad is, even the simpler stuff that was fully comprehensible by common person are also being pushed into the stratosphere, making us dependent on the great creatures that live there.
Now next step is what can we actually learn from these. Any interesting new theorem developed as by-product?
GRH was kind of assumed to be true so this kind of proves the status quo. Like saying you have a proof p! =np. Enormous, but not unexpected.
I think there are proofs it made inside it. Some arithmetic geometry stuff that was all proved to reach the GRH. Overall the proof contains about all of the number theory produced in the last 50 years so I expect a lot to be studied.
>as a Lean library comprising 17 million lines of code.
It looks like the meaning of the word "mathematical proof" changed fundamentally. Mathematics looks like a flattened animal on the asphalt by the CS steamroller.
Correct me if wrong, but didn't lean have a statement that forced the compiler to accept a statement as true with no need to prove it aimed at shortening obvious stuff? If so, are there checks for incorrect usages?
I think it would be useful at this point if mathematicians across various fields worked to come to agreement on which specific sentences in a given formal language correspond to open conjectures. That way the question "Does the Lean proof actually capture the intuitive notion of the problem that we're interested in?" is already answered a priori by having those ready to go.
"We need the human centaur to feed the torment nexus, it's just the best way to build the torment nexus."
Isn't that mathlib?
The TFA described how that exists in MathLib and played an important role in verification.
Maybe I'm missing it then. Can you link me to, e.g., where in the mathlib4 repo P vs NP or Collatz are formalized, and where do I find the names of who reviewed those formalizations?
The closest I've found is Google Deepmind's "Formal Conjectures" repo (https://github.com/google-deepmind/formal-conjectures), but that's not quite what I was thinking of.
As the statement alluded to, that was certainly the case here. The Riemann Hypothesis was formalised and vetted in Lean by humans prior to this proof.
Would be funny if lean has bugs (it's a program after all) and Claude tasked with a hard problem brute forced a vulnerability that allows proving wrong statements.
Would be funny if this was an original thought and not something regurgitated in every thread about this topic
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
TFA addresses this concern
My understanding is that Lean has many independent implementations, and the proof has been verified in several of them. There is very little chance that the same bug would be exploitable across different implementations.
See question 4 here -
https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i...
They only verify Lean proof itself, though it's the same formalization as Mathlib. Still, only human readable proof can actually prove it.
> only human readable proof can actually prove it.
Why?
What if there's a bug in the Lean implementation? How would you detect it?
That's why you use multiple kernels and comparator.
What if a mistake slips peer review?
There could very well be a bug in a single checker, which is why the community is checking the proof across several checker implementations. The article mentioned the proof passing con-leche, which is an external checker with a formal proof of consistency (cannot prove a false statement).
IMO the really interesting bits here are going to be the simplifications that will come from combining ideas from Alpoge and Masullo.
I can imagine the folks at OpenAI have already thrown both papers into the mouths of their leading model.
Until Masullo actually posts a proof instead of the outline of a proof with “fill in proof here” all over the document, there’s not much to work with or reason to get excited about his result. Remember that there have been numerous false proof claims of RH to date.
What are the implications of the Riemann Hypothesis being true?
Why isn’t this acknowledging Masullo’s proof as well?
Masullo's proof didn't come with a Lean formalization.
I think it is an irresponsible lack of context to not cite the proof that has been in progress internally in the mathematics community, however, even if just as a footnote. Anthropic's proof is a slop nuke. A mechanization of Masullo's proof and its surrounding work seems of more practical importance with regards to making idiomatic and reusable formal mathematics than proof-mining Anthropic's proof. Furthermore, I think it's simply insulting not to discuss the work of the mathematical community that both proofs have built upon here...
There are efforts to mechanize Masullo's proof [1]. As of right now, this work is ongoing, so there is nothing to report yet.
With people being as confused as they are (just look at the messages in this thread), I think it's good that TFA is very clear about what it is saying, namely that a specific formal artifact that was posted to the internet is correct, and nothing else. Making value judgements or contrasting the proofs (or proof attempts) is not the point of the post.
(I work at the Lean FRO, but I wasn't involved in the verification effort or the writing of the post)
[1]: https://leanprover.zulipchat.com/#narrow/channel/113488-gene...
I know and I agree with you, but I do think that official communication from the Lean FRO should give appropriate context. A lack of context is equally bad as unclear statements when it comes to communication. Masullo's proof deserves note, even if it is nothing beyond a paragraph stating that Masullo's proof is currently 1) lacking in adequate peer review and 2) undergoing mechanization and verification efforts. Particularly as the AMS announcement explicitly calls for it to be mechanized!
Because it seems to be full of bugs?
It's not a proof. It has very serious gaps.
I don't think this is 100% definitive yet, though
edit: guys that was just my impression when I posted it, it seemed best to state it; do trust the (much more) more qualified people.
I'm not a mathematician, but what do you think is missing?
I'm not a mathematician either, it was just my impression, and I saw that not everything at https://math.mit.edu/~drew/grh had finished running (now several more finished successfully).
How many more checkers should be used before it gets to 100%?
I don't trust anything unless it's checked by someone with a driver's license!
So you'd believe it if a waymo told you? :P
A proof verified by Lean is orders of magnitude more definitive than any peer review, so please tell us how definitive the rest of maths is
I think that highly depends on what's exactly IN the proof and if it accurately tests everything that needs to be tested in the right way, AND have the majority of people agree with all of that first.
The fact that AI is involved with the proofs is probably also controversial.
It's only controversial to the folks who have some moral qualm with it.
Last I heard Alpöge was quite a mathematician in his own right, so saying his using AI is "controversial" is like saying programmers using AI is controversial.
Again, there are some folks super bothered by it for some reason, but most of us normies are just like, "congrats man"
It can’t be both “Claude did this Mathematics is solved” and “The mathematician solved it with claude”. I’d agree it’s probably the latter but there’s a lot of baiting and switching happening where people are talking about the latter but all the press statements say the former.
I think this is now verified to a much higher standard than the vast majority of peer reviewed mathematics of the past decade.