推文

@pirrer · 2026-10-10 19:46

On Oct. 6, @OpenAI released 722 math papers at once. Nearly all came from an internal AI model that isn’t public yet. A day later, OpenAI said about 42% of the main results had Lean proofs, which let a computer check every step. One sign error invalidated three papers, which OpenAI withdrew. Mathematicians are split. Rutgers math chair @AlexKontorovich wrote of one result that if a human did this, “it would be an instant Fields Medal.” NYU math professor Tristan Buckmaster, who has his own dispute with OpenAI, said on CNBC that the release “destroyed the careers” of early-career mathematicians. About four hours before the papers went up, Fields medalist Terence Tao posted a four-part thread that named no company. A breakthrough proof, he wrote, used to bring talks and workshops, draw new people into the field, and end up in textbooks. Now problems often get solved by AI prompters who can’t answer questions about the result. A solved problem can’t be made unsolved, and solutions to open problems are being “harvested at large scale,” leaving entire fields “much less fertile.” One paper claims to prove a conjecture proposed almost 70 years ago by French mathematician Jean-Pierre Serre, who is 100. He told The New York Times that mathematicians enjoy two things, learning and finding new things. AI increases the first and lowers the second, maybe too much, he said, especially for young mathematicians. Tao calls for “Math 2.0”: clear explanations, community building, and new directions should carry more weight. AI can help, he wrote, but not by simply asking it for answers.

曝光 682 · 评论 2 · 点赞 8 · 书签 1 · 曝光/时 132.63466185209222