Post #2373934

2026-05-10 21:59 UTC

Where are the Fields medallists who work outside the topics that @wtgowers and @tao work on, and who are seeing LLMs solve actual problems in their areas of expertise? (Viazovska&#39;s work being autoformalised doesn&#39;t count, it was not somewhat autonomously being improved by AI, being pushed in new directions) https://chadtopaz.com/essays/gowers-response Are they just not being courted by AI companies? Or seeing no good results? We really need to hear analysis of negative results as well. I&#39;ve read Scholze is not a fan, on principle.

Replies (4)

@liuyao@mathstodon.xyz 2026-05-11 01:13

@highergeometer@mathstodon.xyz @wtgowers@mathstodon.xyz @tao@mathstodon.xyz Venkatesh may be someone to look out for. Not sure he has seen advances in his areas yet.

Open ##2525125
@PSL2Z@mathstodon.xyz 2026-05-11 08:18

@highergeometer@mathstodon.xyz Borcherds https://epochai.substack.com/p/ai-math-chat-1-thinking-smarter-not Hairer participates in the First Proof challenge, and his problem wasn't solved in February. A lot of people write about negative results: Litt's takes on this topic are always very sober; benchmarks like FrontierMath are not yet saturated. "Work outside the topics that Tao, a well-known polymath, works on" is an interesting requirement.

Open ##2525126
@cardinal_reinhardt@mastodon.social 2026-05-11 09:02

@highergeometer@mathstodon.xyz i think there are limited negative results because we're still at the point where positive results are remarkable and we acknowledge that. Articles like Gowers' are great for delineating what LLMs are currently capable of without hype. It's difficult to write an article about failed attempts for the same reason it's difficult to write an article about all the times you thought you had a proof of something but there turned out to be a mistake in it...

Open ##2525141
@mathemagical@mathstodon.xyz 2026-05-11 09:37

@highergeometer@mathstodon.xyz @wtgowers@mathstodon.xyz @tao@mathstodon.xyz I find it hard for people with little programming experience to understand how to properly use AI, much less comment on it. I might be too close to AI on this since I've personally been keeping abreast of machine learning since the early days of the recommendation systems in 2011. Do you think these types of posts are more indicative of the state of AI or more the state of the math community's lagging experience with AI? FWIW, my observations after 2 years of lurking in various non-CS/AI communities like math, medicine, etc: - most serious AI practitioners hold their techniques like closely guarded secrets while what permeates socials are either people working at AI companies or AI Hype grifters - a fields introduction to AI follows an eerily predictable pattern and Gowers - Fields Medalists may be geniuses but they're still human and it took me awhile to realize that programming is a completely foreign field to them The biggest mistake I see in all these posts is ignorance that the AI harness not the AI model is what makes or breaks AI usage. You have to program your own problem specific harness around one's own domain problems. This keeps being validated time and time again, most recently with the Firefox team (https://hacks.mozilla.org/2026/05/behind-the-scenes-hardening-firefox/) To use an analogy with cars, it's as if we're at the dawn of invention of automobiles, but everyone is trying to use them as a faster horse-buggy while complaining that cars don't work. The reality is that we need infrastructure and paved roads because of course the car is useless without paved roads, gas stations, etc.

Open ##2525142