I've looked at some of the openai preprints over the last week and am carefully going through a particular one (a proof, not a counterexample, if you must know) with a group of people who have worked on the relevant problem. The impression that is forming is that the people who keep saying ai is bad at writing math and producing slop/going in circles are diluding themselves. This unreleased model is quite good at explaining what it is doing, it is quite good at summarizing what it is doing, it makes choices that are, in some sense, better than those done by a many mathematicians (instead of writing some basic inequalities it says to use Cauchy-Schwartz and that then works...), and it produces an outline of the proof that (after you've looked at the proof for a bit) emphasizes many of the relevant and subtle points. And it crosspolinates ideas quite well (one way to describe it is that it runs the subroutines of different mathematicians to get puts different hurdles, but it is subtler than that..., e.g. it gives a concise summary of the method of a mathematician for a part of the argument that is clear than what that person produced). It is several levels beyong what you get by interacting with fable/astra right now, and just feels like something that is doing this ''without trying too hard''. We'll see how I'll feel in another week or two, there may be big simplifications that it missed (or it may all still collapse, who knows) but right now the openai document is ca. 50 pages and if a group of mathematicians did all these steps, it would probably be much longer.
genuinely feel a bit Cassandra-ish here and worry that if you do not see it happening to something you've thought about for years (even if you are an insider like Dwarkesh) you may not appreciate how far beyong human this is right now.. . . "
No comments:
Post a Comment