techblueberry 2 days ago

I don’t disagree entirely here, but it also comes back to like -

What is a spec.

Can an AI create a better interface than a human? Maybe. Probably even.

But like - interface is a great example where spec and implementation may be the same.

The question becomes how good is your context. Does it all exist well enough for the agent to understand growth patterns, migration strategies. Cost tradeoffs.

If not, then a software engineer guiding architecture may just be a way to use intuition to compress the context created during meetings , from their experience, history, etc. that’s not documented into the spec for the LLM.

It might also be a way to compress weighting. Even if you were able to somehow record every discussion, does that have all the information necessary to weight things properly?

Maybe, but I think we’re still a ways off from that.

And for reason that aren’t exposed I may not want the Optimal solution. If the software engineer came to an architecture intuitively — It would be weird to recommend that a software engineer sort of decompress their intuition so that the LLM could act on the raw business reasoning.

techblueberry 2 days ago

Ok - maybe this is an observation I’m making.

What if humans aren’t smart enough to only feed AI context?

What if we’re incapable of reverse engineering architectural intuitions into well scoped requirements?

I don’t think this is 100% true, but I think it’s a thing I observe.

It’s way easier to prompt an LLM in terms of architectural fundamentals then abstracted business requirements.

  • cyanydeez 2 days ago

    I think you're talking about Wicked Problems: https://en.wikipedia.org/wiki/Wicked_problem

    They're akin to Busy Beaver problems: https://en.wikipedia.org/wiki/Busy_beaver

    They're knowledge, capabilities, etc that arn't just a sum of their parts and like cellular automata, have complex behaviors that are only discoverable via brute force chains.

    So, while brute force is fine in simulate world, eg, navier stokes solutions; in the real world, when something blows up, that's not something "Ooops, I accidently sent the worker into the nuclear reactor because I thought the green alarm meant go".

    That is, basic entropy is what will stop AI from doing much beyond what it's doing currently.

zodwick 2 days ago

100 percent ! most of my time now is spent designing and defining "hills" for claude to climb on