Terence Tao faults OpenAI's 719 AI math proofs for 'proof indigestion'
OpenAI published 722 AI-generated mathematics manuscripts on GitHub on October 6 covering 372 result families and numerous open research problems, withdrawing three the next day after a notation error and its knock-on effects, leaving a public catalogue of 719. Per IT之家, the Institute for Advanced Study-hosted Mathematics and AI Advisory Group (AGMAI) had recommended disclosing model names, prompts, reasoning chains, time and compute costs and stopping tests of hard math problems on proprietary models — yet OpenAI still used proprietary models, only 10 manuscripts include reasoning chains, and about 42% of proofs are unformalized. Mathematician Terence Tao said in an October 7 mathstodon post that he does not oppose AI doing mathematics but rejects making 'fast AI solutions to famous problems' a primary goal or product demo, coining 'proof indigestion' and criticizing the use of problems such as Navier–Stokes as capability benchmarks.
It puts transparency, formal verification and the field's ability to absorb AI math output on the agenda, and shows a leading mathematician publicly questioning speed-and-count measures of understanding.
感谢IT之家网友 不一样的体验 的线索投递！
IT之家 10 月 9 日消息，OpenAI 于 10 月公布 719 份 AI 数学解答，涵盖 372 个数学结果族，涉及数百项开放研究问题，不过发布后被指尚未完全达到“数学与人工智能顾问小组”（AGMAI）提出的标准。
IT之家此前报道，OpenAI 于 10 月 6 日最初在 GitHub 发布 722 份 AI 生成数学手稿总数，涉及 372 个结果族及多项开放研究问题。不过于 10 月 7 日因一处符号错误及其连锁影响撤回 3 份，现公开目录为 719 份。
OpenAI 表示，发布前曾咨询普林斯顿高等研究院主持的“数学与人工智能顾问小组”（AGMAI），并参考其公开建议。然而，从公开信息看，此次发布尚未完全达到该小组提出的标准。
AGMAI 此前建议前沿 AI 实验室停止在专有、外界无法访问的模型上测试高难度数学问题，并要求披露模型名称、提示词、推理链、耗时及计算成本等信息；但 OpenAI 此次仍使用专有模型，且仅 10 份手稿附有模型推理链。
#OpenAI#Terence Tao#AGMAI#Institute for Advanced Study#GitHub#mathstodon