Overview
- OpenAI published 722 manuscripts grouped into 372 result families on GitHub, a set it says came from an unreleased internal model and from roughly 4,000 problems it posed to that system, disclosed Oct. 6.
- The company reported the average published result used about three hours of ChatGPT Pro‑equivalent compute and highlighted claimed advances touching several high‑profile questions including a quasi‑Riemann statement, a Hodge case and a Unique Games claim.
- Only about one‑fifth of the papers include machine‑checked Lean formalizations and OpenAI released abridged chain‑of‑thought summaries for just ten results, and the company warned many unformalized papers could contain issues.
- Mathematicians have demanded the model, prompts and replication records for independent verification and raised unresolved concerns that some AI outputs may draw on unpublished human work while advisory groups and Fields Medal winners pressed for stricter disclosure norms.
- OpenAI says it will fund workshops, add formalizations and record corrections as they arise but institutional checks including peer review and prize adjudication are ongoing and could take months to settle which results stand and how credit is assigned.