What Got Done This Time
Disproving a conjecture and proving a theorem are two different jobs. Finding a counterexample often means searching, within an enormous construction space, for an object satisfying stringent conditions—which happens to suit AI's appetite: it can tirelessly try the bizarre constructions human intuition wouldn't touch. According to the writeup, the counterexample the model produced was verified by human mathematicians and does indeed hold. A conjecture being overturned isn't a bad thing—a false conjecture dying early saves the whole field's time.
In the Context of the Past Two Years
Over the past two years, AI's real track record in math has concentrated on this kind of "search-type" contribution: combinatorial constructions, counterexamples, olympiad problems. They're genuine contributions, but there's still a distance between them and the hype of "AI doing mathematical research"—posing questions worth studying and building new theoretical frameworks remain human territory. The more reasonable picture is: mathematicians have gained a handy exploration tool, like computer-assisted proofs back in the day—first causing discomfort, then becoming routine.
via: Hacker News