OpenAI withdraws three AI-generated math papers due to a sign error

Summary
Dr. Samuel Allen Alexander reviews OpenAI’s update to its openai/math GitHub repository, where three AI-generated mathematics papers were formally withdrawn due to a sign error. He reads through the official changelog documenting the error, the withdrawn papers, and revised manuscripts, and then analyzes the community reactions and debates on Hacker News.

What is shown

• [00:00] Post on X by OpenAI employee Dan Roberts announcing an update to the openai/math repository with 6 new Lean formalizations, 19 modifications, and 3 paper withdrawals, bringing formalized top-line results to ~42%.
• [02:35] Navigation to the openai/math GitHub repository file math/history.md dated October 7, 2026.
• [03:00] Details of the withdrawals section in history.md, specifying that a sign error in Algebraicity of Weil classes on split abelian eightfolds invalidated a stabilization-trace cancellation argument and constructions used in two dependent papers (Algebraicity of Kuga-Satake Correspondences for K3 Surfaces and The rational Hodge conjecture for products of K3 surfaces).
• [04:08] Review of the “Fixes” section of history.md, detailing corrections across 14 manuscripts (e.g., Lipschitz heights and Ashkin-Teller currents, Kähler minimal model programs and abundance, Taming and hypersymplectic deformation) and citation updates across 13 companion papers.
• [05:22] Review of the “Additional Formalizations” section showing 300 out of 719 top-line results formalized (~42%).
• [05:44] Extended review and reading of community comments on Hacker News (“OpenAI withdraws three mathematical results”), covering Lean verification reliability, software engineering methodology applied to mathematics, proof checking, and the impact of AI on mathematical research.

Claims & numbers

• Dan Roberts’ post states the repository update includes 6 new Lean formalizations, 19 modifications, and 3 withdrawals (the presenter reads this directly) [00:12].
• Dan Roberts states the repository now has ~42% of top-line results formalized (300 out of 719) [00:23, 05:32].
• The presenter notes that almost 400 breakthrough results were originally released by OpenAI a few days prior, meaning the three withdrawn papers represent less than 1% of the published results [00:33, 00:43].
• The GitHub repository changelog dated October 7, 2026, confirms 14 manuscripts received proof repairs, corrected statements, and clarified dependencies [04:09].
• Commenter fspeech claims a constraint of “3.5 hours of model effort” per problem was used [07:50].
• The presenter claims Anthropic previously found an elliptic curve of rank 31 that contradicted predictions made in human-written literature [10:10, 24:41].

Notable quotes

• [03:29] “So it isn’t even that there were three different errors. There was just one single error, and ironically of all things, it was a sign error.”
• [14:25] “Now all of a sudden, the proof part is the easy part, and the translation part is the hard part. I never would have predicted that in a billion years.”
• [25:01] “I predict that within the not-too-distant future, AI-produced math papers will be far more reliable than human papers.”

Assessment
This is a commentary and screen-recording review of an official GitHub repository update and related social media/Hacker News discussions. The presenter reads directly from the authentic commit history and forum comments while offering personal analysis and commentary as a mathematician.

Described by gemini-3.8-flash on 2026-10-10 from the video’s audio and frames.