Overview
- OpenAI published the repository on Wednesday, sharing 722 manuscripts that the company says form roughly 372 result‑families and were mostly produced by a single unreleased model.
- OpenAI reported the model was run on about 4,000 tasks and that each included solution used roughly the compute of three hours on ChatGPT Pro, with under 10% of runs judged significant enough for publication.
- The release contains many Lean proof‑assistant formalizations, which make mechanical checking of logical steps possible but do not replace independent peer review of correctness or novelty.
- Questions about provenance and attribution have intensified after earlier claims about a Millennium‑problem‑related solution and an allegation that OpenAI’s work built on prior drafts by Tristan Buckmaster and Levent Alpöge, who is linked to Anthropic.
- OpenAI says it consulted the Institute for Advanced Study’s Advisory Group on Mathematics and Artificial Intelligence and will fund workshops, while wider community validation and formal peer review are still underway and may reshape publication norms and credit practices for machine‑generated research.