A Misalignment of AI in Mathematics
已归档 —— 这条已轮出今日牌堆，完整内容在此保留。
一句话看懂
Twenty-five Fields Medalists published an open letter warning AI problem-solving is misaligned with mathematics.
The signatories say rushed AI proofs threaten attribution, understanding, and math's human chain.
背景
Over recent months, large language models have improved dramatically in mathematics, to the point of solving major outstanding problems across many fields. The declaration emerged from intensive discussions among the 25 mathematicians over the past week, and points to recent incidents where AI systems made claims about long-standing problems, including Millennium Prize problems. The signatories argue that AI companies treating problem-solving as a benchmark is detrimental to mathematics as a science and to its community.
来龙去脉
- Mar 10, 2016In the second game of its match against Lee Sedol, AlphaGo played a move on the fifth line that no professional would have played; it won the game and became known as move 37.
- 2019Lee Sedol retired, saying AI was an entity that could not be defeated.
- 2023A PNAS study graded 5.8 million professional Go moves from 1950 to 2021 with a superhuman engine; play quality was almost flat for sixty-six years, then rose sharply after AlphaGo and kept rising.
- Sep 11, 2026Twenty-five Fields Medalists published an open letter titled "A Severe Misalignment of AI in Mathematics"; Terence Tao posted the full text on his blog "What's New" the same day.
- Sep 12, 2026The statement rose to the top of China's question-and-answer and search boards, setting off debate about the AI industry's benchmark culture; Emad published "Let a billion mathematicians bloom" on Substack.
各方怎么说
- Signatories
- The 25 Fields Medalists say AI companies' push to solve math problems as a benchmark is detrimental to mathematics and its community, and that solving problems is only a tool and proxy metric toward conceptual understanding and insight.
- Signatories
- They warn that mass production of true/false statements at ever-faster pace could destroy fertile ground, that rushed announcements leave no time for writeups, method isolation, or citation, raising severe attribution and plagiarism questions.
- Signatories
- They also say AI offers potential to enhance and accelerate genuine mathematical study, and whether it benefits or harms the field will largely be determined by the humans in control of the technology.
- Commentator
- Emad wrote on Substack: "The Fields Medallists fear AI will break mathematics. I think it may let millions in."
待核实
Rumors that Hodge and BSD proofs exist, with Riemann and P vs NP queued; that FrontierMath Tier 4 is saturated; that Lichtman sees autoformalized math 1,000x away; that an AI system helped yet claims progress on a second Millennium problem, that Andreas Thom got the same non-answer, and that Tworek invokes the dark forest. Also unconfirmed: when an AI company announces progress on such a problem it generates enormous publicity, while later retractions get little attention.