OpenAI has released 722 mathematical manuscripts produced by an unreleased frontier model. The Verge reports that the batch covers 372 result families, which group related papers. Those two counts describe the size of the release, rather than a tally of independently verified solutions.
The release follows OpenAI’s September claim that its model had resolved more than 100 long-standing open problems. The Verge says the Advisory Group on Mathematics and Artificial Intelligence, or AGMAI, describes the new batch as containing solutions to hundreds of open questions. Mathematicians still have to assess the collection’s impact.
What readers can check
In its October 6 announcement, OpenAI says its GitHub repository includes Lean formalizations for many proofs. Lean allows a computer to check mathematical proofs. The company plans to add more formalizations.
The repository also contains ten summaries of model reasoning, attempted-problem statistics and compute estimates. OpenAI puts average compute per result at roughly three hours of ChatGPT Pro thinking. That is a compute comparison, rather than a quoted price for reproducing a result.
Publication and model access follow different schedules
OpenAI says it is working toward releasing the model. Its announcement gives no release date. Researchers can inspect published results before they can use the system that produced them.
The Verge reports that AGMAI’s recommendations ask labs to disclose model names, prompts and compute costs, and to avoid using mathematical releases as marketing. OpenAI says it consulted the group and has revision and citation protocols for this GitHub release.
