Overview
- The papers, which OpenAI posted on GitHub on Tuesday, total 722 manuscripts organized into 372 result families and cover topics from number theory to algebraic geometry.
- OpenAI provided partial materials: 162 reported Lean‑checked main results, ten abridged reasoning summaries, and aggregate compute figures that average about three hours of ChatGPT Pro‑equivalent compute per result.
- Many leading mathematicians protested the dump and an Institute for Advanced Study advisory group urged fuller disclosure; more than two dozen Fields Medal winners warned that rapid, closed testing of frontier problems harms scholarly norms.
- Critics say key artifacts are missing — the internal model, the prompts and full chains of reasoning — leaving independent replication, attribution disputes (including a prior Navier–Stokes complaint) and prize eligibility unresolved.
- OpenAI says it will log corrections as new versions, fund workshops and explore moving the collection to a community repository while work continues across the mathematics community to review, formalize and vet the claims.