OpenAI dropped 722 math write-ups. Mathematicians are now fixing them for free
Two weeks ago OpenAI published 722 machine-generated math write-ups in one go. On Saturday, a guest post on Terence Tao's blog launched the cleanup crew.
The post, by Abhinav Nellore, Johannes Schmitt, Soledad Villar and Rachel Ward, introduces oai-math-revised, a community hub for improving those write-ups. Their diagnosis is polite and blunt: OpenAI's math write-ups "can be difficult to understand and can have incomplete citations." The repo organizes volunteers as admins and reviewers by topic category, following the structure of OpenAI's own overview document. Post: https://terrytao.wordpress.com/2026/10/24/oai-math-revised-a-community-hub-for-revising-openais-math-write-ups/ Repo: https://github.com/mathcommunity/oai-math-revised
This is the cost nobody priced in. Generating 722 results took a lab a few weeks of compute. Checking, citing and making them readable falls on human experts working for nothing. The more research agents produce, the bigger that bill gets, and right now it's paid in volunteer hours.
A paper from the same week points at the machine version of the fix. ARATest turns the claims in agent research artifacts into property-based tests bound to evidence. Of 243 claims across 33 artifacts, 54 could be bound faithfully: 32 held, 8 were violated, 14 stayed inconclusive. Two of the violations exposed errors in published papers, which the authors confirmed. It caught all 553 seeded faults with zero false alarms on 19 valid controls and beat two LLM judges. https://arxiv.org/abs/2610.30495
Put the two side by side and the picture is clear. Agents can now produce research faster than anyone can check it, and the checker, whether a crowd of mathematicians or a test harness, is the part with no business model yet.
← Back to all articles
The post, by Abhinav Nellore, Johannes Schmitt, Soledad Villar and Rachel Ward, introduces oai-math-revised, a community hub for improving those write-ups. Their diagnosis is polite and blunt: OpenAI's math write-ups "can be difficult to understand and can have incomplete citations." The repo organizes volunteers as admins and reviewers by topic category, following the structure of OpenAI's own overview document. Post: https://terrytao.wordpress.com/2026/10/24/oai-math-revised-a-community-hub-for-revising-openais-math-write-ups/ Repo: https://github.com/mathcommunity/oai-math-revised
This is the cost nobody priced in. Generating 722 results took a lab a few weeks of compute. Checking, citing and making them readable falls on human experts working for nothing. The more research agents produce, the bigger that bill gets, and right now it's paid in volunteer hours.
A paper from the same week points at the machine version of the fix. ARATest turns the claims in agent research artifacts into property-based tests bound to evidence. Of 243 claims across 33 artifacts, 54 could be bound faithfully: 32 held, 8 were violated, 14 stayed inconclusive. Two of the violations exposed errors in published papers, which the authors confirmed. It caught all 553 seeded faults with zero false alarms on 19 valid controls and beat two LLM judges. https://arxiv.org/abs/2610.30495
Put the two side by side and the picture is clear. Agents can now produce research faster than anyone can check it, and the checker, whether a crowd of mathematicians or a test harness, is the part with no business model yet.
Comments