Like two or three generations of frontier ago (in the long long ago of last year), I had ChatGPT design a PCB, then Claude a case, then I had China manufacture it on demand and turned it into a real product that shipped to real users. Those chinese factories were mostly staffed by machines, I suspect.
I'm not the only one that had that idea, they're advertising these things as features now.
I don't know what's next, but keep in mind that today's LLM's are the worst they'll ever be.
I hear you - but - I think we are seeing diminishing marginal returns on LLMs already.
They are going to be 'better than what they are today' but I don't see obvious unlocks.
We need more abstractions on top of them.
> "... but keep in mind that today's LLM's are the worst they'll ever be."
That's like saying in 2011: "today's Google search is the worst it will ever be."
Enshitification is coming my friend. We're just now starting to see the kinds of refusals that LLMs will ultimately end up with to protect IP owners and the status quo. We'll look back at this time and marvel that LLMs used to be able to reverse engineer software, or able to create derivative IP works.
What do you think about the fact that gpt sol 6.2 is worse than 6.1 sol (according to code benches and most developers experience) and the new fable 5.2 is not better than opus 5.5 despite costing SIGNIFICANTLY (like 6x more) more p/task?
I wonder if we are seeing them maxing out llm model capabilities now...