Lean Checks Fall Short for AI-Generated Mathematics
Tiger Sniff AI, citing a preprint by Anders Hansen, Fabian Circelli and Alexander Bastounis, reports that part of OpenAI's Navier–Stokes paper differs from the proposition verified by its Lean code. The report also describes Lean 4 kernel vulnerabilities that reportedly allowed false proofs to pass until the Lean team patched them in version 4.33.1.
Passing Lean validation guarantees only that the encoded proposition follows under the checked assumptions; it does not by itself establish that an AI preserved a paper's intended conditions or that the proof-assistant is free of implementation flaws.

今年9月,Anthropic宣布,AI用11天把费马大定理的证明写成1300万行Lean代码。几天后,OpenAI又宣称用AI解决了纳维–斯托克斯难题。两项成果都附有通过Lean检查的代码。但数学家随后发现,OpenAI文字论文中的部分中间结论,与Lean代码实际验证的内容并不一致。Lean被视为数学的“机器......
本文来自微信公众号: 返朴 ,作者:任平生
今年9月,Anthropic宣布,AI用11天把费马大定理的证明写成1300万行Lean代码。几天后,OpenAI又宣称用AI解决了纳维–斯托克斯难题。两项成果都附有通过Lean检查的代码。但数学家随后发现,OpenAI文字论文中的部分中间结论,与Lean代码实际验证的内容并不一致。