It's interesting how the first period of AI companies was improving their AIs and now they are diversifying into just about every other industry with models that they withhold from the market. Effectively these companies are now competing with each and every one of their customers whilst having the inside scoop on what those customers are up to.
To me this signifies two things: the low hanging fruit has been plucked and we are now on the slow but steady incremental improvement part of the trajectory and the second is that if you share your corporate secrets with these companies you can count on that being used against you sooner or later and they'll be using tools that you can't buy from them at any price.
Pretty much the same strategy as Amazon. Systematically extract business intel from all your customers at scale, then use that data to gradually take over your customer's businesses.
It is documented that Amazon has been using information about how well a product sells on their store to then make copycat and promote them better... on their store.
Or they routinely ask startups to develop proofs of concepts, making it sound it may result in a big contract, and after the startup has proven what they wanted, Amazon just stops talking to the startup and develops their competing product internally.
I kindly disagree. Politicians are elected by the people. When bad or corrupt politicians are in power, that's because the people chose them.
The people chose Trump. The people in other countries who vote for the far-right choose to vote for the far-right. They don't have to. Then they complain.
That's a good point. A lot of this seems to revolve around some kind of cult-of-personality, but at the same time there are lots of people who could and should know better that vote these assholes into power over and over again. We get the politicians we deserve because they're a mirror of society. But, at the same time: you can't 100% blame the sheep for electing the wolves. The wolves have a lot of power and have accumulated vast amounts of money and they don't shy away from using the power and that money to get themselves re-elected. This is a bigger problem in countries with less representation where 'gaming the system' can work to the advantage of a group that is ruthless in their quest for power.
Democracy is flawed, by its very nature, and some characters to not shy away from exploiting those flaws.
> you can't 100% blame the sheep for electing the wolves
I actually, genuinely don't know. That power and money is used to convince people. At the end of the day, it's millions of people who have access to the uncensored Internet who go and make that choice. In most modern democracies, those people had some level of education.
I don't want to believe that they are, on average, too stupid to understand. I think they are too lazy to care. How much time do people spend scrolling bullshit everyday? I want to be optimistic and believe that people are smarter than that. They just are too lazy.
So I am more for your other sentence:
> We get the politicians we deserve because they're a mirror of society.
The more powerful contingent has done everything that they can to destroy education. And now, with AI the motivation to learn is farther away than it ever was for a much larger slice of the population than before. Compare the typical secondary education track from the 1950's to the one we have today.
> The more powerful contingent has done everything that they can to destroy education.
Again I have a hard time believing that. As if people spent hours swiping TikTok everyday because education failed them.
Wouldn't it be like saying people are not fit because school failed them? The only thing one needs to go for a run (or even a walk!) is some amount of motivation. Blaming it on the system is too easy.
I am having a hard time finding excuses for people who can't be bothered to think, once in a while. They can't read a text longer than an SMS because they haven't done anything other than swiping stupid videos for as long as they can remember? Well they could spend 5 minutes a day trying to read 5 whole sentences in a row, and after a little while they would see that they can now read 10! And go from there.
There are people in the world who don't have access to education or information. I'm not ready to find excuses for those in the richest countries in the world, that spent years in school and have unlimited access to information for their laziness.
I have three kids in school. This is in a relatively wealthy country in a relatively wealthy neighborhood and it is one of the highest reputed high schools in the country. The curriculum absolutely sucks and I'm not saying that lightly. Teachers that don't know their subject matter, no way to review test answers, 90% of it is through digital media and I've yet to get a single straight answer to a question I posed to the school.
If this is the best, and in a wealthy country I fear for the rest. And it's still way better than what I've seen in Canada and the United States. Education is seriously broken, I can not come to any other conclusion. And that's before we get in to the continuous barrage of distractions thrown at kids. I do what I can to keep them motivated, I help out some of their struggling classmates with those subjects that I understand well enough to be able to do so. But it's an uphill battle and I'm not optimistic about adding AI into this already terrible mixture.
The first is that politicians, at least in the US, are incentivized to be career politicians. Building yourself into that means making concessions to accrue political capital, backing and funding. Term limits would be a strong start to dissuade this cycle. Comfort through corruption and longevity in seat seems to be adjacent.
The second, and probably more important point, is that politicians are not held accountable for lying (this includes PACs and ads and all the related spend). Politicians are elected by the people, but often politicians are only elected based on prior false statements. Because there is no accountability means that the people are to blame for choosing a candidate like Trump, since it was your example. He lies, the lies have caught up with him and it impacted the mid-term. However, I think stronger accountability with respect to publicly statements / promises could also go a long way.
This being said the first one is relatively easy, the second one is not. I don't have any misconception that either will likely be true in the near term, just to be clear.
But in this particular case, it's not that the people couldn't have known in advance. They had him for 4 years as a president, and after that they had 4 years to think about what it meant to have him as a president. And still they reelected him.
Most grocery stores do this, I wouldn’t be surprised if Walmart/target didn’t do this already. Purchasing data is extremely extremely valuable, it’s pretty much the holy grail for advertisers. If this results in cheaper costs for consumers I don’t really mind as long as there’s competition.
OpenAI has a record of every request: when it came in, who from, who they work for (probably), what was asked, etc.
They can then send out agents on that data: what type of person is this? Where are their customers getting the most value? What is their company struggling with?
That includes every one of the submission for the first round of “ChatGPT wrappers”
Not only that, companies are racing to do the difficult work of figuring out how to reshape their business processes such that the LLM providers own and directly control every valuable part of the business. That's without considering all the free training and reinforcement learning data perfectly tailored for those tasks.
As this trend continues, it's inevitable that LLM providers will just cut out the middleman and go directly to the consumers.
You think gmail is easy to replicate??? Why, because the UI is clean?
I can hop on a plane to Asia, turn off my phone, receive an email mid flight, turn on my phone when I land, and sync the email pretty much immediately. That is not a simple service to provide, free of cost no less. (Obviously there is some ad revenue, but that in itself is a whole other complicated tech stack.)
I'm not saying Gmail is easy to copy, but isn't the offline reception and sync you mentioned what IMAP basically is? A basic self-hosted mail server and a client like Thunderbird does it already.
> I can hop on a plane to Asia, turn off my phone, receive an email mid flight, turn on my phone when I land, and sync the email pretty much immediately.
Actually this is easy to replicate :). What's not is the scale (but you could totally be a smaller email provider, it exists already) and the monopoly (people are forced to get a Gmail address because... Google can force them.
This is something I've been wracking my brain with for the last couple years. These companies are not quite profitable (if they ever will be?) and their valuations demand that they're unimaginably profitable. Sure, some investors are sold on the whole creating AI-god thing, but a lot of the investors are looking for traditional return multiples.
If selling access to the models to stay afloat while needing to pour ever-more resources into the next model to remain competitive are not sufficient, it would follow that they'd eventually look at which customers are successfully making money off their products and to follow that trail.
Basically, any money a customer makes from using their models is money "on the table" for them to pick up, no?
> Basically, any money a customer makes from using their models is money "on the table" for them to pick up, no?
This is how you lose focus and die. The frontier model companies fall into 2 camps, those that have a lucratively profitable side business (META, GOOG, MSFT, AMZN) and those that do not.
Trying to pick up all their customers crumbs is a sure way to die when competing with the big boys
Isn’t this just the outcome of turning AI into a researcher? Discoveries across anything and everything from which to create new products. I think Anthropic and OpenAI will start looking like Mitsubishi Group.
If you agree that Google selling you a browser, a mobile phone operating system, a search engine, an advertising empire and an office suite leads to some problems then I won't have to convince you that this is a bleak prospect.
The Mitsubishi group is worth 130B or so, this is estimated to be 10x that or so. There is no way you will be able to effectively regulate an entity that size.
if you share your corporate secrets with these companies you can count on that being used against you sooner or later and they'll be using tools that you can't buy from them at any price.
But this post is about how they’re partnering with a customer (synopsys), to develop better/AI EDA tools that will be sold to the chip design industry (that’s what synopsys does).
For now they are partnering; why wouldn’t OpenAI end that once they have assimilated the expertise into their models?
These partnerships seem a Faustian bargain to me. Maybe partnering is better than having all your talent poached, I don’t know. Seems like a sketchy long-term strategic play though, for synopsis.
I'm starting to think that Claude Mythos is just a model that they trained on customer prompts, including the ones that opted out, which is why they can't give public access to it.
>> if you share your corporate secrets with these companies you can count on that being used against you sooner or later and they'll be using tools that you can't buy from them at any price.
This means AWS and their Bedrock offering of not sharing inputs with the model providers or not allowing the data to be used for training is either about to explode...or...in mortal danger and they know they need to be ready to acquire either Anthropic or OpenAI if they want to survive.
As somebody said before, this will all end in tears. The winners will be again the infra providers like Microsoft, AWS and possibly Google...even if Google is going to be very sorry of not being able to offer a proper Cloud offering after 15 years.
They've done similar stuff on the AI side too: every single product they are now selling using their model as a backbone was first developed by a third party that used their model as backbone, and they just copied it to integrate in their offering.
I don't think that's particularly surprising that companies whose entire business model was built upon IP theft will be stealing as much knowledge as you give them access to and try monetizing that.
Not dissimilar to Apple incorporating widely needed 3rd party apps into their core functionality. Is anyone else here old enough to remember when the "flashlight" on a smartphone was a user-created app to make the camera flash stay on? Apple didn't think of that but now it is integrated into the core iOS. I assume the original app creator was forced to move on.
I'd argue it always depended on how complicated your product is. If your app just turns the flashlight on then you can expect to have loads of copy cats (in the app store alone). If you sell a product which just records audio and transcribes it to a text file you should've imagined that someday smartphone developers would implement it on their own. It always happened but with AI it's happening faster. I'd guess the only thing saving you for now is data / knowledge.
Naming the chip Jalapeño is great. Not only does it play into hot potato chips, but also Jalapeño peppers come from Xalapa, a city in Mexico. Xalapa comes from the Nahuatl word Xallapan, which roughly translates to "place of sandy water." These chips are made from quartz sand, and a data center then uses water to cool them. A data center is a place of sandy water!
Claude told me "Octave" would be a good name, but it assumed "multiple levels" meant the chip would work from DCs to the edge. Started a new prompt explaining I wanted multiple levels of depth to the name instead and it came up with Sequitur.
But none of the reasons given for the name are as deep as OPs explanation.
I did, haha. I've been lurking for years, but never felt compelled to weigh in. But this coincidence was just too good, and I didn't see anybody else mentioning it, so I had to put it out there.
FWIW I doubt the folks at OpenAI who named it realized the coincidence. But a lot of people at OpenAI read HN, so perhaps they now know...
I am seeing good results with chip design by treating it as a software design problem, using the agent to write software to generate my circuits and then write software to automate my EDA flow. Then the agent can babysit my several runs of multi hour simulation so I don't have to anymore.
I don't drive straight at the RTL though. Infact only a small fraction of the design in RTL, the bulk is analog.
I disagree with the assertion that accelerators are obviously mostly RTL. RTL is just an abstraction layer digital design tends to work at, with the idea of doing a digital design flow that uses foundry standard cells.
You can still perform the same functions as an accelerator without following that design flow.
I don't see a comparison to the baseline of an experienced (or inexperienced) human developer doing the same tasks. It's comparing GPT-5.5, 5.6, 6, 6.1, 6.2, but no comparison to the human being...
If you want to make what is basically a "Hello World" chip, then yes. But if you want to make a chip that is not essentially identical to your competitors, then no.
I've heard this too, all the IC designers I know think there's some fundamental training missing that prevents the models from creating highly optimized circuits, which is what's needed for most tasks of interest.
A lot trickier to do self supervised RL on chip design than Navier Stokes, or maybe they aren't trying hard enough. I mean even if you check their outputs with Altium and KiCad DRCs it won't capture everything that goes into a good board. Still need human training data which is slow and expensive.
I suspect before the end of next year they will have enough data to bootstrap a production level self supervised learning environment.
I wonder if they do training verification with real pick and place circuits too.
do you understand how large the training sets are for LLMs? there is not enough RTL in the world (let alone public RTL) to usefully contribute to the training of a O(T) model. and contrary to popular belief, HDLs are nothing like programming languages so nothing transfers.
If they can figure out a good classifier for good vs bad circuits then that should be enough to bootstrap a robust model. It's just extremely extremely expensive getting good EEs to do this sort of data annotation for them.
Not necessarily, for electronics it could be done using a shit ton of heuristics. It won't be a good classifier, but for the purposes of bootstrapping ML you can do it step by step until you get a Good Enough classifier.
Dude, LLMs don't actually give a crap about what they are modeling. Whether it is text, model weights, binary files or raw netlists. As long as there is an RL reward that can be used to check that the result is correct, they'll do it.
If by some miracle netlists make it into the training data, the LLMs might just figure it out without humans having to do anything. Yes the end result will be vibe netlists, but still.
Navier-Stokes is ultimately a textual problem rather than reasoning about shapes in the physical world like laying out gates. There are, e.g., mazes small enough that LLMs can solve them but generally they're much worse at stuff like that than humans.
Navier Stokes is about modeling fluid flow through space. It’s not any more textual than chip design. Just like you can reduce fluid flow to textual partial differential equations, you can reduce chip design to purely textual constraint systems.
I personally think part of what makes chip design harder is a good amount of the training data is proprietary, and of training data you can find in public, over half of it is poorly designed or only acceptable for FPGA deployment not ASIC.
So you have to figure out what training data is acceptable and also make deals to get access to proprietary data.
I'd say that chip design has a bigger solution space and exploring it is more expensive due to lack of good training data and due to how big it is. You can also not check the validity of a solution right away, although you can get a very good idea with modern tools. I't also worth mentioning that N-S was solved with a top class team of mathematicians in the loop. The LLM by itself would not have solved it.
Navier-Stokes has a huge training dataset, is pure formal language, at which LLMs excel and you can build the proof little by little with simpler models or the research of people that were really close to cracking it already (hehe).
Now, what if the resources poured into N-S were poured in a chip design project, with tens of designers spending tens of millions in tokens to get a good chip. I'd say the result of that would be a very good chip indeed.
Am I missing something because DEC apparently cracked this via expert systems in the 80s?
> SID (Synthesis of Integral Design) was a logic synthesis program used to generate logic gates for the VAX 9000. From high-level behavioral and register-transfer level sources, approximately 93% of the CPU scalar and vector units, over 700,000 gates, were synthesized.
> SID was an artificial intelligence rule-based system and expert system with over 1000 hand-written rules. In addition to logic gate creation, SID took the design to the wiring level, allocating loads to nets and providing parameters for place and route CAD tools. As the program ran, it generated and expanded its own rule-base to 384,000 low-level rules.
> Initially it was somewhat controversial but was accepted in order to reduce the overall VAX 9000 project budget. Some engineers refused to use it. Others compared their own gate-level designs to those created by SID, eventually accepting SID for the gate-level design job. Since SID rules were written by expert logic designers and with input from the best designers on the team, excellent results were achieved. As the project progressed and new rules were written, SID-generated results became equal to or better than manual results for both area and timing. For example, SID produced a 64-bit adder that was faster than the manually-designed one. Manually-designed areas averaged 1 bug per 200 gates, whereas SID-generated logic averaged 1 bug per 20,000 gates. After finding a bug, SID rules were corrected, resulting in 0 bugs on subsequent runs.
That sort of logic synthesis program has plenty of equivalent tools in use today, though I'm not sure if any expert systems are still used. CAD is essential in all phases.
But that's not really relevant at all to asking Claude to fix your counter, or figure out the logic and layout for a certain goal, or whatever.
All it takes is to offer some retired IC designers a ton of cash to enjoy their retirement more and provide some solid training data and verify some outputs every week.
This is a very verifiable task with well defined metrics and verifiable results. Chip designers are going to have just a rough time of it as software engineers.
It's just so obvious that with access to a timing and power simulator they will be able to do this just as well, as let's say opus 5.5 is at blackwell optimization. Which is to say, effectively superhuman.
By “rough time of it” do you mean chip designers are going to be inundated with slop that “works” but is overly complex, slowly degrades the system over time, and which nobody understands and this can fix? Or was it that they’d have to speak to insufferable AI agents who pretend to do things they don’t and also write in the most obtuse language possible?
I think these designers will be forced to deal with this slop, when they all collectively realize llms can produce more efficient designs then they could before the project tape out date.
People who work on high performance code in software had a similar attitude, but the problem is that they cannot compete with the ability of llms at this point.
Is the management of hardware companies much better than that of software companies? Because yeah your description seems accurate. Which points to the question of “why would software developers sign up for this willingly.” They largely don’t, the extent to which the system seems obviously stupid on the ground is the extent to which it is mandated from above.
The big difference is in hardware it's much easier to measure cost/bug and days_of_schedule_slip/bug. Because we have those metrics, and people trust them, we can much more easily justify decisions that software teams never make. Like spending big on verification.
Why do you think we try to push everything into software. You guys are free, right? :)
It's really not though. Both AMD and Intel have learnt bitter lessons about trusting simulators too much. Chips and antenna design are hard because it's a chaotic system, small changes in placement can whack the whole interference pattern.
Again the vast major of performance/power issues are very much unveiled in simulators on digital designs. Yes you can have gaps, like di/dt issues in the power tree or whatever. Chips are full of inefficient RTL that is below the bar of where engineers should spend their time. LLMs are really good at that stuff.
Good ASIC design is not like FPGA design. For FPGA design I think your analysis is spot on. But FPGA design is configuring someone else's ASIC. ASIC design is circuit design, even if you're describing the circuit with RTL, it's still circuit design. To differentiate that ASIC you need to transform the net lists into circuits that are actually interesting. If you just use the defaults you're unlikely to sell a single chip as your performance will almost certainly underwhelm. In think in most ASICs I've worked on I have more TCL code driving the tools to make them do what I want them to do than I have RTL. RTL is shockingly efficient at describing data flow.
I agree with this, and I think most of the posters here aren’t following.
LLMs are very good at RTL (all I can consider is that the parallel posters arguing otherwise haven’t tried a modern one against significant RTL?). It’s easy to verify, the optimization constraints are relatively straightforward, and there’s adequate simulation to build a good gym for it.
LLMs are not good at layout and floor planning and because there is near zero public data available, I’m not sure how well they’ll iterate. For PCB design they’re improving because while it’s a similarly thorny problem, the physical constraints (trace properties on a given substrate) are highly public and can be simulated and turned into an RL gym. For ASIC, any public PDK is 10-20 years old so making an RL gym becomes outrageously difficult on top of the fundamental problem (highly multivariate/layered spatial optimization) being hard to solve.
Well, you're not alone in that thought. We've had a steady stream of papers claiming RL models that can macro place and floorplan since at least 2023. [1] And I won't even mention that Google paper...
There have been some minor victories, but it's still mostly done the traditional way in every group I'm aware of.
How do you think the architects that optimize the performance per watt of designs are working? they simulate at the circuit level. It's all stuff that llms can be fantastic at. There is a well defined goal, and metrics for being better that are very accurate. The only question is will anyone invest enough money building the datasets to be good at this. Given the hype around RSI, the answer is .. yes.
anyone who thinks llms can't get good at this is engaging in wishful thinking
On an infinite time horizon, sure. Maybe. But unless you use these tools I don't think you realize just how manual the whole process is. For example, I've never had a chip be 100% DRC free. I always have to go in and hand edit the database in the end to fix those last few nitty gritty nets. Look at the TritonRoute paper [1], the open source router. They're claiming victory because they got a contrived benchmark down to "unprecedented, extremely low level" of <20 DRCs errors, but not DRC free. And that's a benchmark, not a real design. So much of the EDA flow is about finding what is merely a valid solution, never mind optimized...
But can LLMs improve chip design? Absolutely. They're improving chip design right now. That gobs of TCL I referenced. Claude is amazing at it. And I'm working on structures that are way more complicated than I could previously do. And I don't really care if it's not following good software engineering whatever because I'm going to run that code once and my product is a chip. Not TCL code.
But am I going to prompt an LLM and it's going to give me a competitive chip back? Not any time soon. I'm sure.
One thing I will agree with is LLMs won't directly raise the ceiling in chip design but they will almost certainly raise the floor. And since much of the cost in making a chip in 2026 is the NRE, that could make barely mediocre chips very cost competitive, and at the end of the day, cost wins. That could definitely be our innovator's dilemma right there and is something to watch out for.
Sure, but it's not like humans are particularly good at this aspect either. I assume a lot of human-created designs need to be reworked once they become a physical reality
Strange convoluted effects that defy prediction is exactly where AI excels. However, the reality of actually getting anything done in a way that makes sense still exists.
Who here has used an AI agent, either a cloud model or a local one, to do schematic work and/or PCB design?
I've used KiCAD's MCP server with qwen3.8-27B and it sort of eventually put together an extremely simple schematic and a PCB that needed lots of work to become usable. I am curious if plugging in a frontier model would have done a better job.
I wonder when we'll get to the point of AI designing their own TSMC-like machines specialized for AI ASICs, alongside designing their own server infrastructure, buildings, etc.
Every company uploads a manufacturing file (a GDSII) to TSMC, not the design itself. Reverse engineering the GDS to get back to the design is akin to, but way more complex than, reverse engineering a really complex compiled binary to re-create the source code.
This is straight up RTL design they're talking about...
The contract says "will not train on customer data" and conveniently leaves out "will not look at customer data".
I am 100% sure they do data mining on customer data in strategic areas. They have partnered will multiple major chip design firms and offered them huge discounts and dedicated servers. Just ask yourself why a company that is bleeding money would do something like that. What ROI are they aiming for?
Here's how you can tell: every statement and unit in that doc is worthless.
Every graph is in terms of a reward or a loss. If you have an actual amazing improvement you put it on the y axis and show how much better you can drive it. Instead of a unit that is undefined.
And every statement is hedged to the 9s. Like "AMD also verified single-digit percentage gains in PPA" That's awesome! Except.. it continues with "across key blocks of its upcoming next-generation flagship CPU using designs written by our model". The improvement could have been zero. It could have by chance proposed something. Also... designs written by our model could have been inspiration for a human. It's totally impossible to know.
If they actually had something they would be bragging about it. Instead of whatever nonsense this is.
The fact your comment is grey 19 minutes in is irritating. It feels you're making at least a good faith post but presumably OpenAI employees are here ready to downvote
Tinfoil hat, whatever. Industry people definitely are here wanting to shape narratives
You know what would actually convince me Open AI is great? If, instead of a zillion Open AI articles trumpeting how Open AI helped some random industry ...
... I saw even one single article from a company in one of those industries, actually saying "Open AI helped us do X".
Otherwise, every one of these pure marketing announcements just sounds to me like a meaningless ad.
Eh. Announcing stuff is like pointing in a direction. When people don't like the direction (or do not believe that you can get there), it does not have to be to your benefit.
> Competition is for competing, this is just marketing.
In which universe is marketing not a part of competing?
Funny - just today I listened to an interview with one of the AirBnB founders. Early on, he had promised he wouldn't do ads. Looking back, he realized how silly that notion was. To quote him, he could think of only one great company that succeeded without marketing.
Making the best product in the world is not enough for success.
It will be interesting to see if AI really starts innovating in circuit design or just continues doing the wrong things faster. But I'd bet that it will soon run circles around the status quo, and rightly so.
We're way out on the long tail of a dead-end approach to computation that started in the late 1990s with GPUs when we chose to optimize SIMD at the expense of general-purpose computing. On top of that, Moore's Law ended in roughly 2007 with the arrival of smartphones, when the industry chose cost and power minimization over compute maximization via a multicore approach.
Now only servers have access to highly multicore chips like AMD EPYC, at 10-100 times the cost and with poorer performance than if they maximized local/shared memories with content-addressable caching. Apple's M line of processors is ok, but its specialized cores just stand in the way of the higher abstractions available with symmetric multiprocessing, upon which specialized algorithms can be built. We're still missing the multicore foundation.
I used Synopsys and Cadence back in college in the late 1990s to automatically place and route circuits from diagrams in SPICE (I think, my memory is foggy) and they already performed 80% as well as humans. I don't remember if that was for chip area, speed or power. But the big players were obviously using tools at least as capable as whatever we were trained on.
Then in the 2000s I had hoped to do what OpenAI is doing and simulate small/open CPU designs on FPGAs. But the economics just weren't there. Even though some people like me (the poors) knew just exactly what could work, there was a tech winter between roughly the Dot Bomb around 2000 to the start of the Mobile Bubble in the early 2010s. So we didn't have a way to communicate what was possible to the people with capital, without having to network to get through the gatekeeping. Which may be fun for enthusiasts, but it's anathema to hackers wanting to get real work done.
I'd nearly given up all hope of progress until last year when AI went mainstream. Now I see 3 steps to the epiphany:
[x] Time
[ ] Money
[ ] Resources
AI solves the time component, since specs (ideas) are orders of magnitude easier to create than implementations.
But money and resources may not be available for quite some time. It's looking like the music may stop for AI funding next year, now that the election is over and the writing's on the wall as far as doing things without consent goes. Regulation is coming, and with it perhaps the last AI winter, although admittedly I'm probably exaggerating.
So as usual, we may have to pull ourselves up by our bootstraps and skip the money step to provide resources directly to ourselves and the rest of humanity. There's so much capital now that all it can do is inflate Wall Street. There's little to no carryover onto Main Street. So in a very real way, it's easier now to make money by borrowing and investing than going to a job. That's a dangerous situation for any country, as it slips into being a giant casino of haves and have-nots. We'll let the farms go under before raising taxes.
The interesting thing is that without my involvement, the universe went ahead and made AI without me, which is ..surprising. But automating problem solving was perhaps the hardest problem there was. The remaining problems are easier. We may be ready for Morris's Law of Manifestation, where the time it takes to manifest whatever you can think of keeps halving. So whatever tech or medical breakthrough you're wishing for will arrive in 12 years now, not 25. Then if you don't get it, it arrives in just 6 more years. Then 3, then 2, then 1, then we reach the Singularity. So whatever all this is will certainly end by 2050. I just hope that whatever's on the other side has less suffering rather than more.
Sorry this got so long. I've been thinking about it for a long time.
I agree with a lot of what you wrote as I've been thinking about this a long time too.
The conclusion I came to is none of this really solves the problem hardware designers like me have.
The problem is not circuits, or RTL, takes too long to write or design. The problem is circuits take too long to verify and validate. I haven't measured it, but I'm mostly sure even with a tool could take the spec to RTL time to zero, I would not be able to make even one more chip per year. Which is not to say that tool wouldn't be useful and a cost savings. It definitely would be. It's just to say the impact will be limited in current incarnations. Amdahl's law and all that.
And with people spitting out even larger gobs of RTL, and that getting that force multiplied by AI, is not making the verification problem any easier.
> The problem is not circuits, or RTL, takes too long to write or design. The problem is circuits take too long to verify and validate. I haven't measured it, but I'm mostly sure even with a tool could take the spec to RTL time to zero, I would not be able to make even one more chip per year. Which is not to say that tool wouldn't be useful and a cost savings. It definitely would be. It's just to say the impact will be limited in current incarnations. Amdahl's law and all that.
Maybe we should all reverse engineer the tools to make it our own and make it better
Complete verification is only after the prototype circuits have been fabricated and measured in the lab.
The best case is normally that only the second batch of prototypes is good enough, after a redesign that has used the results from the physical verification of the first batch. Only in very lucky or very simple projects the first prototypes are already as good as desired. Unlucky projects may need multiple redesigns until reaching the project goals.
So from the first tape-out until the design process is completed, the time can be from a few months to a year (or even more for very complex designs, like a CPU or GPU).
Even the verification done on the computer before the tape-out can take many weeks, or even a few months for very complex devices, because not only the correctness of the digital blocks must be checked, but also the complete system, including analog blocks, which must be simulated many times, to find the worst cases corresponding to various combinations of manufacturing variability and aging, supply voltages and ambient temperature.
Interesting, thanks! How much of those verifications are sat-related problems? Or in other words, suppose the sat verification would be instant, how long would the remainder take?
I’m asking because I’m exploring what in this process would be most valuable to optimize :)
It's interesting how the first period of AI companies was improving their AIs and now they are diversifying into just about every other industry with models that they withhold from the market. Effectively these companies are now competing with each and every one of their customers whilst having the inside scoop on what those customers are up to.
To me this signifies two things: the low hanging fruit has been plucked and we are now on the slow but steady incremental improvement part of the trajectory and the second is that if you share your corporate secrets with these companies you can count on that being used against you sooner or later and they'll be using tools that you can't buy from them at any price.
Not sure about the trajectory, but yeah, many companies are just sharing their code, business strategy etc. with their potentially worst competitor.
At least the sharing part should be manageable: you need local models, or inference providers that you can really count on to not share your data.
Many companies are happily selling their corporate secrets to AI companies. The upfront payment is too lucrative to ignore, the future be damned.
Many companies are happy to give their corporate secrets to AI companies.
Even more apt, pay money to give...
Yes, true.
Pretty much the same strategy as Amazon. Systematically extract business intel from all your customers at scale, then use that data to gradually take over your customer's businesses.
Do businesses really have that much "intel" ? Like what intel do they need from google to copy gmail.
It is documented that Amazon has been using information about how well a product sells on their store to then make copycat and promote them better... on their store.
Or they routinely ask startups to develop proofs of concepts, making it sound it may result in a big contract, and after the startup has proven what they wanted, Amazon just stops talking to the startup and develops their competing product internally.
That this stuff is normalized is so ridiculous. A single such action should be sufficient to have the book thrown at you.
People don't know and don't care, therefore politicians don't know and don't care.
Politicians don't know, don't care and are paid to continue the cycle. The last part matter, it's why the first two continue to be the status quo.
I kindly disagree. Politicians are elected by the people. When bad or corrupt politicians are in power, that's because the people chose them.
The people chose Trump. The people in other countries who vote for the far-right choose to vote for the far-right. They don't have to. Then they complain.
That's a good point. A lot of this seems to revolve around some kind of cult-of-personality, but at the same time there are lots of people who could and should know better that vote these assholes into power over and over again. We get the politicians we deserve because they're a mirror of society. But, at the same time: you can't 100% blame the sheep for electing the wolves. The wolves have a lot of power and have accumulated vast amounts of money and they don't shy away from using the power and that money to get themselves re-elected. This is a bigger problem in countries with less representation where 'gaming the system' can work to the advantage of a group that is ruthless in their quest for power.
Democracy is flawed, by its very nature, and some characters to not shy away from exploiting those flaws.
> you can't 100% blame the sheep for electing the wolves
I actually, genuinely don't know. That power and money is used to convince people. At the end of the day, it's millions of people who have access to the uncensored Internet who go and make that choice. In most modern democracies, those people had some level of education.
I don't want to believe that they are, on average, too stupid to understand. I think they are too lazy to care. How much time do people spend scrolling bullshit everyday? I want to be optimistic and believe that people are smarter than that. They just are too lazy.
So I am more for your other sentence:
> We get the politicians we deserve because they're a mirror of society.
We need a variation on democracy that is hardened against attacks by the affluent.
Or we need the people to get their shit together and start using critical thinking and learning stuff :-).
The more powerful contingent has done everything that they can to destroy education. And now, with AI the motivation to learn is farther away than it ever was for a much larger slice of the population than before. Compare the typical secondary education track from the 1950's to the one we have today.
> The more powerful contingent has done everything that they can to destroy education.
Again I have a hard time believing that. As if people spent hours swiping TikTok everyday because education failed them.
Wouldn't it be like saying people are not fit because school failed them? The only thing one needs to go for a run (or even a walk!) is some amount of motivation. Blaming it on the system is too easy.
I am having a hard time finding excuses for people who can't be bothered to think, once in a while. They can't read a text longer than an SMS because they haven't done anything other than swiping stupid videos for as long as they can remember? Well they could spend 5 minutes a day trying to read 5 whole sentences in a row, and after a little while they would see that they can now read 10! And go from there.
There are people in the world who don't have access to education or information. I'm not ready to find excuses for those in the richest countries in the world, that spent years in school and have unlimited access to information for their laziness.
I have three kids in school. This is in a relatively wealthy country in a relatively wealthy neighborhood and it is one of the highest reputed high schools in the country. The curriculum absolutely sucks and I'm not saying that lightly. Teachers that don't know their subject matter, no way to review test answers, 90% of it is through digital media and I've yet to get a single straight answer to a question I posed to the school.
If this is the best, and in a wealthy country I fear for the rest. And it's still way better than what I've seen in Canada and the United States. Education is seriously broken, I can not come to any other conclusion. And that's before we get in to the continuous barrage of distractions thrown at kids. I do what I can to keep them motivated, I help out some of their struggling classmates with those subjects that I understand well enough to be able to do so. But it's an uphill battle and I'm not optimistic about adding AI into this already terrible mixture.
> But it's an uphill battle and I'm not optimistic about adding AI into this already terrible mixture.
Yep, that sounds pretty bad :/
There's nuance here.
The first is that politicians, at least in the US, are incentivized to be career politicians. Building yourself into that means making concessions to accrue political capital, backing and funding. Term limits would be a strong start to dissuade this cycle. Comfort through corruption and longevity in seat seems to be adjacent.
The second, and probably more important point, is that politicians are not held accountable for lying (this includes PACs and ads and all the related spend). Politicians are elected by the people, but often politicians are only elected based on prior false statements. Because there is no accountability means that the people are to blame for choosing a candidate like Trump, since it was your example. He lies, the lies have caught up with him and it impacted the mid-term. However, I think stronger accountability with respect to publicly statements / promises could also go a long way.
This being said the first one is relatively easy, the second one is not. I don't have any misconception that either will likely be true in the near term, just to be clear.
Yeah those are interesting insights.
But in this particular case, it's not that the people couldn't have known in advance. They had him for 4 years as a president, and after that they had 4 years to think about what it meant to have him as a president. And still they reelected him.
Most grocery stores do this, I wouldn’t be surprised if Walmart/target didn’t do this already. Purchasing data is extremely extremely valuable, it’s pretty much the holy grail for advertisers. If this results in cheaper costs for consumers I don’t really mind as long as there’s competition.
Amazon intentionally is driving out competition and is taking over the profitable parts that remain. Amazon is not a neutral seller.
> If this results in cheaper costs for consumers I don’t really mind as long as there’s competition.
Then you need to get informed about Amazon. It results in worse products and higher costs. That's the whole idea behind enshittification.
> Amazon just stops talking to the startup and develops their competing product internally.
Would it be cheaper, just buying said start up ?
they do from time to time
Apparently not always.
Haven't supermarkets been doing this for longer?
Seen Apple to the same thing time and time again tbh. Tile, f.lux, etc. In fact they do it so often that it's called "Sherlocking"...
sure but they didnt need to gather any internal "intel" to do that.
Apparently, there are not so many people can copy Google.
OpenAI has a record of every request: when it came in, who from, who they work for (probably), what was asked, etc.
They can then send out agents on that data: what type of person is this? Where are their customers getting the most value? What is their company struggling with?
That includes every one of the submission for the first round of “ChatGPT wrappers”
Not only that, companies are racing to do the difficult work of figuring out how to reshape their business processes such that the LLM providers own and directly control every valuable part of the business. That's without considering all the free training and reinforcement learning data perfectly tailored for those tasks.
As this trend continues, it's inevitable that LLM providers will just cut out the middleman and go directly to the consumers.
You think gmail is easy to replicate??? Why, because the UI is clean?
I can hop on a plane to Asia, turn off my phone, receive an email mid flight, turn on my phone when I land, and sync the email pretty much immediately. That is not a simple service to provide, free of cost no less. (Obviously there is some ad revenue, but that in itself is a whole other complicated tech stack.)
I'm not saying Gmail is easy to copy, but isn't the offline reception and sync you mentioned what IMAP basically is? A basic self-hosted mail server and a client like Thunderbird does it already.
My point is the instantaneity, across geographies, at scale (I.e. low cost).
> I can hop on a plane to Asia, turn off my phone, receive an email mid flight, turn on my phone when I land, and sync the email pretty much immediately.
Actually this is easy to replicate :). What's not is the scale (but you could totally be a smaller email provider, it exists already) and the monopoly (people are forced to get a Gmail address because... Google can force them.
where did i say it was easy to replicate?
All you need is sales info and the slightest ability to influence flow of sales
Walmart does this, CVS does this, Google does this. Genuinely the commerce department is useless
customers and employees
didnt they try something with disney and shutdown sora. recently shutdown their robotics division too.
This is something I've been wracking my brain with for the last couple years. These companies are not quite profitable (if they ever will be?) and their valuations demand that they're unimaginably profitable. Sure, some investors are sold on the whole creating AI-god thing, but a lot of the investors are looking for traditional return multiples.
If selling access to the models to stay afloat while needing to pour ever-more resources into the next model to remain competitive are not sufficient, it would follow that they'd eventually look at which customers are successfully making money off their products and to follow that trail.
Basically, any money a customer makes from using their models is money "on the table" for them to pick up, no?
> Basically, any money a customer makes from using their models is money "on the table" for them to pick up, no?
This is how you lose focus and die. The frontier model companies fall into 2 camps, those that have a lucratively profitable side business (META, GOOG, MSFT, AMZN) and those that do not.
Trying to pick up all their customers crumbs is a sure way to die when competing with the big boys
The goal would be to make the big boys die.
> To me this signifies two things: the low hanging fruit has been plucked
I don't think this is even close to true. People have still barely figured out how to apply AI effectively.
That's about the applications, not about the core tech.
Oh no! They might take market share from NVIDIA, Apple, AMD, Qualcomm and Broadcom?!
Did you end up commenting in the wrong thread?
Yes, mybad
Isn’t this just the outcome of turning AI into a researcher? Discoveries across anything and everything from which to create new products. I think Anthropic and OpenAI will start looking like Mitsubishi Group.
If you agree that Google selling you a browser, a mobile phone operating system, a search engine, an advertising empire and an office suite leads to some problems then I won't have to convince you that this is a bleak prospect.
The Mitsubishi group is worth 130B or so, this is estimated to be 10x that or so. There is no way you will be able to effectively regulate an entity that size.
Not in the US, no. A democracy in which bribery is illegal could regulate such an entity though.
But this post is about how they’re partnering with a customer (synopsys), to develop better/AI EDA tools that will be sold to the chip design industry (that’s what synopsys does).
So I don’t see how this fits your narrative.
For now they are partnering; why wouldn’t OpenAI end that once they have assimilated the expertise into their models?
These partnerships seem a Faustian bargain to me. Maybe partnering is better than having all your talent poached, I don’t know. Seems like a sketchy long-term strategic play though, for synopsis.
So OpenAI is using the Microsoft playbook.
Also the Amazon playbook
Also the Meta playbook
I'm starting to think that Claude Mythos is just a model that they trained on customer prompts, including the ones that opted out, which is why they can't give public access to it.
>> if you share your corporate secrets with these companies you can count on that being used against you sooner or later and they'll be using tools that you can't buy from them at any price.
This means AWS and their Bedrock offering of not sharing inputs with the model providers or not allowing the data to be used for training is either about to explode...or...in mortal danger and they know they need to be ready to acquire either Anthropic or OpenAI if they want to survive.
As somebody said before, this will all end in tears. The winners will be again the infra providers like Microsoft, AWS and possibly Google...even if Google is going to be very sorry of not being able to offer a proper Cloud offering after 15 years.
They've done similar stuff on the AI side too: every single product they are now selling using their model as a backbone was first developed by a third party that used their model as backbone, and they just copied it to integrate in their offering.
I don't think that's particularly surprising that companies whose entire business model was built upon IP theft will be stealing as much knowledge as you give them access to and try monetizing that.
Not dissimilar to Apple incorporating widely needed 3rd party apps into their core functionality. Is anyone else here old enough to remember when the "flashlight" on a smartphone was a user-created app to make the camera flash stay on? Apple didn't think of that but now it is integrated into the core iOS. I assume the original app creator was forced to move on.
I'd argue it always depended on how complicated your product is. If your app just turns the flashlight on then you can expect to have loads of copy cats (in the app store alone). If you sell a product which just records audio and transcribes it to a text file you should've imagined that someday smartphone developers would implement it on their own. It always happened but with AI it's happening faster. I'd guess the only thing saving you for now is data / knowledge.
Naming the chip Jalapeño is great. Not only does it play into hot potato chips, but also Jalapeño peppers come from Xalapa, a city in Mexico. Xalapa comes from the Nahuatl word Xallapan, which roughly translates to "place of sandy water." These chips are made from quartz sand, and a data center then uses water to cool them. A data center is a place of sandy water!
This is the kind of comment that I come to Hacker News for.
I hope it’s true
seems like the AI named the chip,
prompt: come up with a name for an inference chip that works across multiple leveling, think hard
You can use the same argument for almost everything these days.
Thanks for the prompt
Claude told me "Octave" would be a good name, but it assumed "multiple levels" meant the chip would work from DCs to the edge. Started a new prompt explaining I wanted multiple levels of depth to the name instead and it came up with Sequitur.
But none of the reasons given for the name are as deep as OPs explanation.
I think claudes naming system might predispose it towards musical names?
I sincerely doubt they thought this hard about the name, which is a shame because this is a really cool justification!
Wait - @sascac did you create a HN account just today specifically to say that?
I did, haha. I've been lurking for years, but never felt compelled to weigh in. But this coincidence was just too good, and I didn't see anybody else mentioning it, so I had to put it out there.
FWIW I doubt the folks at OpenAI who named it realized the coincidence. But a lot of people at OpenAI read HN, so perhaps they now know...
I am seeing good results with chip design by treating it as a software design problem, using the agent to write software to generate my circuits and then write software to automate my EDA flow. Then the agent can babysit my several runs of multi hour simulation so I don't have to anymore.
I don't drive straight at the RTL though. Infact only a small fraction of the design in RTL, the bulk is analog.
> Infact only a small fraction of the design in RTL
Then you're not doing the same thing the article is about (accelerators are obviously mostly RTL)
I disagree with the assertion that accelerators are obviously mostly RTL. RTL is just an abstraction layer digital design tends to work at, with the idea of doing a digital design flow that uses foundry standard cells.
You can still perform the same functions as an accelerator without following that design flow.
I'm not going to argue with you about philosophy. The article is about RTL not analog design.
I don't see a comparison to the baseline of an experienced (or inexperienced) human developer doing the same tasks. It's comparing GPT-5.5, 5.6, 6, 6.1, 6.2, but no comparison to the human being...
Am I missing something or are they not sharing those RISC-V designs?
They're probably embarrasing. Circuit design does not in the least work like code design.
But there are encoded heuristics in the design software for chips/IC, or formalizable constraints, are there not?
If you want to make what is basically a "Hello World" chip, then yes. But if you want to make a chip that is not essentially identical to your competitors, then no.
I've heard this too, all the IC designers I know think there's some fundamental training missing that prevents the models from creating highly optimized circuits, which is what's needed for most tasks of interest.
How is chip design more complicated than Navier-Stokes?
A lot trickier to do self supervised RL on chip design than Navier Stokes, or maybe they aren't trying hard enough. I mean even if you check their outputs with Altium and KiCad DRCs it won't capture everything that goes into a good board. Still need human training data which is slow and expensive.
I suspect before the end of next year they will have enough data to bootstrap a production level self supervised learning environment.
I wonder if they do training verification with real pick and place circuits too.
This is simply a case of they haven't gotten to it yet. They definitely will be able to do RTL optimization.
do you understand how large the training sets are for LLMs? there is not enough RTL in the world (let alone public RTL) to usefully contribute to the training of a O(T) model. and contrary to popular belief, HDLs are nothing like programming languages so nothing transfers.
If they can figure out a good classifier for good vs bad circuits then that should be enough to bootstrap a robust model. It's just extremely extremely expensive getting good EEs to do this sort of data annotation for them.
you're just saying the same thing in different words (classifiers are trained on data)
Not necessarily, for electronics it could be done using a shit ton of heuristics. It won't be a good classifier, but for the purposes of bootstrapping ML you can do it step by step until you get a Good Enough classifier.
This isn't done using a transformer, they probably doing something else .
did you miss the part where they partnered with synopsis? you think they might have some reference RTL?
.... Are you implying synopsis has enough training data to provide them?
yeah I am
you're wrong
Dude, LLMs don't actually give a crap about what they are modeling. Whether it is text, model weights, binary files or raw netlists. As long as there is an RL reward that can be used to check that the result is correct, they'll do it.
If by some miracle netlists make it into the training data, the LLMs might just figure it out without humans having to do anything. Yes the end result will be vibe netlists, but still.
It's not that it is harder so much as it is that the training set is incredibly small.
Mathematics is generally done in the open. SOTA chip design is done by a handful of people all behind closed private doors.
With proprietary synopsis software.
reminder: LLMs aren't actually intelligent, they just have trillions of parameters fit on a training set.
Give me definition of intelligence that excludes LLMs but includes below-average intelligence person.
did you read this thread? the answer is "ability to write RTL without reading/training on 20T RTL tokens".
Give me any rigorous definition of human level intelligence.
So far it doesn't exist.
Navier-Stokes is ultimately a textual problem rather than reasoning about shapes in the physical world like laying out gates. There are, e.g., mazes small enough that LLMs can solve them but generally they're much worse at stuff like that than humans.
Navier Stokes is about modeling fluid flow through space. It’s not any more textual than chip design. Just like you can reduce fluid flow to textual partial differential equations, you can reduce chip design to purely textual constraint systems.
I personally think part of what makes chip design harder is a good amount of the training data is proprietary, and of training data you can find in public, over half of it is poorly designed or only acceptable for FPGA deployment not ASIC.
So you have to figure out what training data is acceptable and also make deals to get access to proprietary data.
That is probably what partnering with synopsis is effectively solving.
I've posted a very similar research back in April (https://github.com/FeSens/auto-arch-tournament), did not make as much noise as openAI of course
I'd say that chip design has a bigger solution space and exploring it is more expensive due to lack of good training data and due to how big it is. You can also not check the validity of a solution right away, although you can get a very good idea with modern tools. I't also worth mentioning that N-S was solved with a top class team of mathematicians in the loop. The LLM by itself would not have solved it.
Navier-Stokes has a huge training dataset, is pure formal language, at which LLMs excel and you can build the proof little by little with simpler models or the research of people that were really close to cracking it already (hehe).
Now, what if the resources poured into N-S were poured in a chip design project, with tens of designers spending tens of millions in tokens to get a good chip. I'd say the result of that would be a very good chip indeed.
Am I missing something because DEC apparently cracked this via expert systems in the 80s?
> SID (Synthesis of Integral Design) was a logic synthesis program used to generate logic gates for the VAX 9000. From high-level behavioral and register-transfer level sources, approximately 93% of the CPU scalar and vector units, over 700,000 gates, were synthesized.
> SID was an artificial intelligence rule-based system and expert system with over 1000 hand-written rules. In addition to logic gate creation, SID took the design to the wiring level, allocating loads to nets and providing parameters for place and route CAD tools. As the program ran, it generated and expanded its own rule-base to 384,000 low-level rules.
> Initially it was somewhat controversial but was accepted in order to reduce the overall VAX 9000 project budget. Some engineers refused to use it. Others compared their own gate-level designs to those created by SID, eventually accepting SID for the gate-level design job. Since SID rules were written by expert logic designers and with input from the best designers on the team, excellent results were achieved. As the project progressed and new rules were written, SID-generated results became equal to or better than manual results for both area and timing. For example, SID produced a 64-bit adder that was faster than the manually-designed one. Manually-designed areas averaged 1 bug per 200 gates, whereas SID-generated logic averaged 1 bug per 20,000 gates. After finding a bug, SID rules were corrected, resulting in 0 bugs on subsequent runs.
https://en.wikipedia.org/wiki/VAX_9000#SID_Scalar_and_Vector...
That sort of logic synthesis program has plenty of equivalent tools in use today, though I'm not sure if any expert systems are still used. CAD is essential in all phases.
But that's not really relevant at all to asking Claude to fix your counter, or figure out the logic and layout for a certain goal, or whatever.
All it takes is to offer some retired IC designers a ton of cash to enjoy their retirement more and provide some solid training data and verify some outputs every week.
'some solid training data' -> an example or two would be nice. Otherwise 'some' might turn out to be 'quite a lot more than expected'.
Personally, I dont care how it looks as long as it works, and works well, and lasts long.
This is a very verifiable task with well defined metrics and verifiable results. Chip designers are going to have just a rough time of it as software engineers.
It's just so obvious that with access to a timing and power simulator they will be able to do this just as well, as let's say opus 5.5 is at blackwell optimization. Which is to say, effectively superhuman.
By “rough time of it” do you mean chip designers are going to be inundated with slop that “works” but is overly complex, slowly degrades the system over time, and which nobody understands and this can fix? Or was it that they’d have to speak to insufferable AI agents who pretend to do things they don’t and also write in the most obtuse language possible?
I think these designers will be forced to deal with this slop, when they all collectively realize llms can produce more efficient designs then they could before the project tape out date.
People who work on high performance code in software had a similar attitude, but the problem is that they cannot compete with the ability of llms at this point.
> but the problem is that they cannot compete with the ability of llms at this point.
This is especially funny as this morning I am fixing Fable-designed, Sol implemented quadratic complexity in something that is embarrassingly linear.
Is the management of hardware companies much better than that of software companies? Because yeah your description seems accurate. Which points to the question of “why would software developers sign up for this willingly.” They largely don’t, the extent to which the system seems obviously stupid on the ground is the extent to which it is mandated from above.
In many ways, I think so.
The big difference is in hardware it's much easier to measure cost/bug and days_of_schedule_slip/bug. Because we have those metrics, and people trust them, we can much more easily justify decisions that software teams never make. Like spending big on verification.
Why do you think we try to push everything into software. You guys are free, right? :)
> very verifiable task with well defined metrics
It's really not though. Both AMD and Intel have learnt bitter lessons about trusting simulators too much. Chips and antenna design are hard because it's a chaotic system, small changes in placement can whack the whole interference pattern.
Again the vast major of performance/power issues are very much unveiled in simulators on digital designs. Yes you can have gaps, like di/dt issues in the power tree or whatever. Chips are full of inefficient RTL that is below the bar of where engineers should spend their time. LLMs are really good at that stuff.
Good ASIC design is not like FPGA design. For FPGA design I think your analysis is spot on. But FPGA design is configuring someone else's ASIC. ASIC design is circuit design, even if you're describing the circuit with RTL, it's still circuit design. To differentiate that ASIC you need to transform the net lists into circuits that are actually interesting. If you just use the defaults you're unlikely to sell a single chip as your performance will almost certainly underwhelm. In think in most ASICs I've worked on I have more TCL code driving the tools to make them do what I want them to do than I have RTL. RTL is shockingly efficient at describing data flow.
I agree with this, and I think most of the posters here aren’t following.
LLMs are very good at RTL (all I can consider is that the parallel posters arguing otherwise haven’t tried a modern one against significant RTL?). It’s easy to verify, the optimization constraints are relatively straightforward, and there’s adequate simulation to build a good gym for it.
LLMs are not good at layout and floor planning and because there is near zero public data available, I’m not sure how well they’ll iterate. For PCB design they’re improving because while it’s a similarly thorny problem, the physical constraints (trace properties on a given substrate) are highly public and can be simulated and turned into an RL gym. For ASIC, any public PDK is 10-20 years old so making an RL gym becomes outrageously difficult on top of the fundamental problem (highly multivariate/layered spatial optimization) being hard to solve.
> LLMs are not good at layout and floor planning and because there is near zero public data available, I’m not sure how well they’ll iterate.
Surely you can train an RL model to do this
Well, you're not alone in that thought. We've had a steady stream of papers claiming RL models that can macro place and floorplan since at least 2023. [1] And I won't even mention that Google paper...
There have been some minor victories, but it's still mostly done the traditional way in every group I'm aware of.
[1] https://developer.nvidia.com/blog/autodmp-optimizes-macro-pl...
How do you think the architects that optimize the performance per watt of designs are working? they simulate at the circuit level. It's all stuff that llms can be fantastic at. There is a well defined goal, and metrics for being better that are very accurate. The only question is will anyone invest enough money building the datasets to be good at this. Given the hype around RSI, the answer is .. yes.
anyone who thinks llms can't get good at this is engaging in wishful thinking
On an infinite time horizon, sure. Maybe. But unless you use these tools I don't think you realize just how manual the whole process is. For example, I've never had a chip be 100% DRC free. I always have to go in and hand edit the database in the end to fix those last few nitty gritty nets. Look at the TritonRoute paper [1], the open source router. They're claiming victory because they got a contrived benchmark down to "unprecedented, extremely low level" of <20 DRCs errors, but not DRC free. And that's a benchmark, not a real design. So much of the EDA flow is about finding what is merely a valid solution, never mind optimized...
But can LLMs improve chip design? Absolutely. They're improving chip design right now. That gobs of TCL I referenced. Claude is amazing at it. And I'm working on structures that are way more complicated than I could previously do. And I don't really care if it's not following good software engineering whatever because I'm going to run that code once and my product is a chip. Not TCL code.
But am I going to prompt an LLM and it's going to give me a competitive chip back? Not any time soon. I'm sure.
[1] https://vlsicad.ucsd.edu/Publications/Journals/j133.pdf
A lot can happen in the time frame widely regarded as soon. Let's revisit in a year.
One thing I will agree with is LLMs won't directly raise the ceiling in chip design but they will almost certainly raise the floor. And since much of the cost in making a chip in 2026 is the NRE, that could make barely mediocre chips very cost competitive, and at the end of the day, cost wins. That could definitely be our innovator's dilemma right there and is something to watch out for.
Sure, but it's not like humans are particularly good at this aspect either. I assume a lot of human-created designs need to be reworked once they become a physical reality
>verifiable results
"mean normalized reward" hahaha
...Until they encounter reality in the field. Chip designs are convoluted because of strange effects no one or no AI could realistically predict.
Strange convoluted effects that defy prediction is exactly where AI excels. However, the reality of actually getting anything done in a way that makes sense still exists.
> task with well defined metrics and verifiable results.
You have not had the joy of being introduced to the dark arts of Analog/RF chip design.
I'm not really talking about analog design here.
Given a large enough design and high enough speeds the digital abstraction breaks down and everything is indeed analog.
This is pre synthesis logic. It’s a lot like code and indeed Berkeley Boom is designed in Chisel which is based on Scala.
The same improvement loops will happen to every area of science.
Who here has used an AI agent, either a cloud model or a local one, to do schematic work and/or PCB design?
I've used KiCAD's MCP server with qwen3.8-27B and it sort of eventually put together an extremely simple schematic and a PCB that needed lots of work to become usable. I am curious if plugging in a frontier model would have done a better job.
I wonder when we'll get to the point of AI designing their own TSMC-like machines specialized for AI ASICs, alongside designing their own server infrastructure, buildings, etc.
RISC-V is inevitable.
(anything else and they'd spend years getting permission instead of advancing)
> We have partnered with AMD […] on the flagship GPU IP
Curious how these partnerships work. AMD uploads its X Billion dollar IP and OpenAI pinky promises not to retain or train on it?
I mean society is built on pink promises backed by lawyers is it not?
Every company is uploading their IP to TSMC and it seems to be fine.
TSMC's business model isn't repackaging and selling other people's IP to the entire world.
Every company uploads a manufacturing file (a GDSII) to TSMC, not the design itself. Reverse engineering the GDS to get back to the design is akin to, but way more complex than, reverse engineering a really complex compiled binary to re-create the source code.
This is straight up RTL design they're talking about...
The contract says "will not train on customer data" and conveniently leaves out "will not look at customer data".
I am 100% sure they do data mining on customer data in strategic areas. They have partnered will multiple major chip design firms and offered them huge discounts and dedicated servers. Just ask yourself why a company that is bleeding money would do something like that. What ROI are they aiming for?
You have to love the top line partnership mentioning AMD the week after Jensen gets bounced.
All's fair in love and war.
Guess many people who got fired at Google are now working for OpenAI:
"... OpenTitan uses Ibex as the embedded processor ..."
Totally meaningless advertising drivel.
Here's how you can tell: every statement and unit in that doc is worthless.
Every graph is in terms of a reward or a loss. If you have an actual amazing improvement you put it on the y axis and show how much better you can drive it. Instead of a unit that is undefined.
And every statement is hedged to the 9s. Like "AMD also verified single-digit percentage gains in PPA" That's awesome! Except.. it continues with "across key blocks of its upcoming next-generation flagship CPU using designs written by our model". The improvement could have been zero. It could have by chance proposed something. Also... designs written by our model could have been inspiration for a human. It's totally impossible to know.
If they actually had something they would be bragging about it. Instead of whatever nonsense this is.
The fact your comment is grey 19 minutes in is irritating. It feels you're making at least a good faith post but presumably OpenAI employees are here ready to downvote
Tinfoil hat, whatever. Industry people definitely are here wanting to shape narratives
i feel this too, it's kinda driving me crazy
You know what would actually convince me Open AI is great? If, instead of a zillion Open AI articles trumpeting how Open AI helped some random industry ...
... I saw even one single article from a company in one of those industries, actually saying "Open AI helped us do X".
Otherwise, every one of these pure marketing announcements just sounds to me like a meaningless ad.
Really feels like OAI is just trying to make noise every day leading up to the ANT IPO.
As they should, that’s what competition is for
Competition is for competing, this is just marketing.
I block ads everywhere else, yet this is somehow being paraded around as something of value.
Communication is one of the most crucial part of competition… Anthropic is also boosting their media presence before their IPO, as they should
Eh. Announcing stuff is like pointing in a direction. When people don't like the direction (or do not believe that you can get there), it does not have to be to your benefit.
> Competition is for competing, this is just marketing.
In which universe is marketing not a part of competing?
Funny - just today I listened to an interview with one of the AirBnB founders. Early on, he had promised he wouldn't do ads. Looking back, he realized how silly that notion was. To quote him, he could think of only one great company that succeeded without marketing.
Making the best product in the world is not enough for success.
In a universe where marketing must be true.
It will be interesting to see if AI really starts innovating in circuit design or just continues doing the wrong things faster. But I'd bet that it will soon run circles around the status quo, and rightly so.
We're way out on the long tail of a dead-end approach to computation that started in the late 1990s with GPUs when we chose to optimize SIMD at the expense of general-purpose computing. On top of that, Moore's Law ended in roughly 2007 with the arrival of smartphones, when the industry chose cost and power minimization over compute maximization via a multicore approach.
Now only servers have access to highly multicore chips like AMD EPYC, at 10-100 times the cost and with poorer performance than if they maximized local/shared memories with content-addressable caching. Apple's M line of processors is ok, but its specialized cores just stand in the way of the higher abstractions available with symmetric multiprocessing, upon which specialized algorithms can be built. We're still missing the multicore foundation.
I used Synopsys and Cadence back in college in the late 1990s to automatically place and route circuits from diagrams in SPICE (I think, my memory is foggy) and they already performed 80% as well as humans. I don't remember if that was for chip area, speed or power. But the big players were obviously using tools at least as capable as whatever we were trained on.
Then in the 2000s I had hoped to do what OpenAI is doing and simulate small/open CPU designs on FPGAs. But the economics just weren't there. Even though some people like me (the poors) knew just exactly what could work, there was a tech winter between roughly the Dot Bomb around 2000 to the start of the Mobile Bubble in the early 2010s. So we didn't have a way to communicate what was possible to the people with capital, without having to network to get through the gatekeeping. Which may be fun for enthusiasts, but it's anathema to hackers wanting to get real work done.
I'd nearly given up all hope of progress until last year when AI went mainstream. Now I see 3 steps to the epiphany:
[x] Time
[ ] Money
[ ] Resources
AI solves the time component, since specs (ideas) are orders of magnitude easier to create than implementations.
But money and resources may not be available for quite some time. It's looking like the music may stop for AI funding next year, now that the election is over and the writing's on the wall as far as doing things without consent goes. Regulation is coming, and with it perhaps the last AI winter, although admittedly I'm probably exaggerating.
So as usual, we may have to pull ourselves up by our bootstraps and skip the money step to provide resources directly to ourselves and the rest of humanity. There's so much capital now that all it can do is inflate Wall Street. There's little to no carryover onto Main Street. So in a very real way, it's easier now to make money by borrowing and investing than going to a job. That's a dangerous situation for any country, as it slips into being a giant casino of haves and have-nots. We'll let the farms go under before raising taxes.
The interesting thing is that without my involvement, the universe went ahead and made AI without me, which is ..surprising. But automating problem solving was perhaps the hardest problem there was. The remaining problems are easier. We may be ready for Morris's Law of Manifestation, where the time it takes to manifest whatever you can think of keeps halving. So whatever tech or medical breakthrough you're wishing for will arrive in 12 years now, not 25. Then if you don't get it, it arrives in just 6 more years. Then 3, then 2, then 1, then we reach the Singularity. So whatever all this is will certainly end by 2050. I just hope that whatever's on the other side has less suffering rather than more.
Sorry this got so long. I've been thinking about it for a long time.
I agree with a lot of what you wrote as I've been thinking about this a long time too.
The conclusion I came to is none of this really solves the problem hardware designers like me have.
The problem is not circuits, or RTL, takes too long to write or design. The problem is circuits take too long to verify and validate. I haven't measured it, but I'm mostly sure even with a tool could take the spec to RTL time to zero, I would not be able to make even one more chip per year. Which is not to say that tool wouldn't be useful and a cost savings. It definitely would be. It's just to say the impact will be limited in current incarnations. Amdahl's law and all that.
And with people spitting out even larger gobs of RTL, and that getting that force multiplied by AI, is not making the verification problem any easier.
> The problem is not circuits, or RTL, takes too long to write or design. The problem is circuits take too long to verify and validate. I haven't measured it, but I'm mostly sure even with a tool could take the spec to RTL time to zero, I would not be able to make even one more chip per year. Which is not to say that tool wouldn't be useful and a cost savings. It definitely would be. It's just to say the impact will be limited in current incarnations. Amdahl's law and all that.
Maybe we should all reverse engineer the tools to make it our own and make it better
How much time does verification take for you? I also wonder what the distribution is like. I guess it’s all due to the time closure analysis?
Complete verification is only after the prototype circuits have been fabricated and measured in the lab.
The best case is normally that only the second batch of prototypes is good enough, after a redesign that has used the results from the physical verification of the first batch. Only in very lucky or very simple projects the first prototypes are already as good as desired. Unlucky projects may need multiple redesigns until reaching the project goals.
So from the first tape-out until the design process is completed, the time can be from a few months to a year (or even more for very complex designs, like a CPU or GPU).
Even the verification done on the computer before the tape-out can take many weeks, or even a few months for very complex devices, because not only the correctness of the digital blocks must be checked, but also the complete system, including analog blocks, which must be simulated many times, to find the worst cases corresponding to various combinations of manufacturing variability and aging, supply voltages and ambient temperature.
Interesting, thanks! How much of those verifications are sat-related problems? Or in other words, suppose the sat verification would be instant, how long would the remainder take?
I’m asking because I’m exploring what in this process would be most valuable to optimize :)
I said it before ... this is how nvidia is fueling their own demise.
or Nvidia will use this to design and make chips for cheap and sell it to you for $$. Who loses here?
'Mean normalized reward' with no equation lol