Remix.run Logo
▲ doctorpangloss 6 hours ago

i don't think that people are wrong to use the words "brute force" to describe the models - swarms of computers - writing 10,000 proposals for Lean proofs, implementing them all, and seeing which compile and continuing until some advancement of the solution, a 1,000,000:1 ratio of work : moving towards the goal, as the superhuman step needed to finish these proofs. Most forces in our world are superhuman.

▲darepublic 2 hours ago | parent [-]

It's spaghetti at the wall. Lean is crucial to clear away the failed brute force exploration. Just like in chess long ago. .brute forcing lines.. you needed an evaluation. Maybe an unpopular sentiment but took how many millions on tokens? So a lot of it is just generator to try different ideas in this complicated forest and the lean compilability as some kind of compass

▲IshKebab an hour ago | parent [-]

If you actually watch AI writing Lean it is very very far from brute force. Recent models are extremely good at Lean. When I was trying to get them to teach me it, I would paste my failed tactic into a non-agent Opus 5.5 and just say "why doesn't this work" and it could easily tell me the exact error message I was seeing without even running a Lean compiler. Obviously this is on very simple code but still, hardly brute force.

I'm working on a quite expansive project and most of its time actually seems to be spent avoiding accidental complexity blowup in the tactics it writes, rather than any difficulty writing the actual proofs. Obviously it's not Navier-Stokes, but it's still super complex stuff. Easily PhD level and way beyond my skill.