Thread by @littmath

One piece of info that seems important to me in terms of forecasting usefulness of new AI models for mathematics: did the gold-medal-winning models, which did not solve IMO problem 6, submit incorrect answers for it? 🧵

Why this is important: it is very hard/time-intensive to check correctness of English-language mathematical proofs. (See below for a thread on this topic.)

https://x.com/littmath/status/1874559283859501567

A tool that can solve hard math problems but also produces incorrect answers to problems it can't solve could end up being a net productivity sink--or even worse, if it fools the user, lead to incorrect claims.

One worry I have is that AI tools will develop their ability to produce hard-to-check, convincing mathematical prose more rapidly than their ability to produce correct answers. If so I suspect we will soon see a deluge of persuasive-looking, incorrect results.

I've recently noticed an uptick in crank papers in my area submitted to arXiv -- 7 out of 9 papers with "Hodge conjecture" in the title or abstract since June 15 have been nonsense. I think almost all of these were written with LLM aid.

Currently it's quite easy for me to check these papers are nonsense (a matter of a few minutes) but as capabilities improve, this will become more difficult. It's quite possible we will soon no longer be able to trust a randomly-chosen math paper on arXiv.

Of course (auto)formalization will help to address this--but in my area any sort of formalization is, I think, pretty far away.

If these models did not submit a wrong answer to P6, I think that's a pretty bullish sign for the near-term future of AI-for-math. If they did... /fin