The Pulse
OpenAI Publishes 719 Manuscripts of AI-Generated Math
OpenAI published 719 math manuscripts, grouped into 372 families, from an unreleased internal model. The repository includes Lean proof formalizations, details about the model’s problem-solving process and warnings that some results still n
AI.info Team · Written with AI ·
OpenAI’s October 6 release puts 719 manuscripts online
OpenAI published a collection of AI-generated mathematical work on October 6, placing 719 manuscripts in a public GitHub repository. The papers are grouped into 372 families of related results, ranging across mathematical disciplines and theoretical computer science. The company says an unreleased internal model produced the work.
The collection is not a set of 719 independently confirmed breakthroughs. A family can include a main result, related arguments, consequences or alternative proofs, and OpenAI says the material sits at different stages of verification. Some results also build on earlier outputs from its models.
The repository README warns that some results without formal proofs “could have issues,” and says OpenAI will try to fix them. It also records revisions as new versions while keeping earlier versions accessible, giving readers a way to track changes to the manuscripts.
About 4,000 problems went into the evaluation
OpenAI says it posed approximately 4,000 problems to the model during the evaluation. Aggregating the outputs into result families and manuscripts, and requiring an appropriate level of significance, yielded the published catalogue. The company says the vast majority followed the same process with its internal model.
“The average result used the equivalent compute of roughly three hours of ChatGPT Pro thinking,” OpenAI wrote. The company identifies exceptions, including work on a zero-free region for the Riemann zeta function and a proof concerning the Hodge conjecture for CM abelian varieties. It says a write-up about one zeta-function result was edited by a human for readability.
Lean formalizations cover some, not all, results
OpenAI’s repository includes formalizations in Lean, a programming language used to express mathematical proofs in a form a computer can check. The company says it will add more formalizations as it obtains them. The repository reports that about 42% of its top-line results have been formalized; the remaining results do not all have that form of computer-checkable proof.
That distinction matters because the release is a research collection, not a certification that every claim is correct or significant. OpenAI’s post says it is publishing additional information about the process, including 10 summaries of model reasoning, estimates of compute and statistics on attempted problems. The model itself remains unreleased.
Advisory group says publication is only a first step
The Advisory Group on Mathematics and Artificial Intelligence, an independent group hosted at the Institute for Advanced Study, responded to the release on October 6. It said its advisory role should not be taken as a judgment of the results’ impact or an endorsement of how OpenAI produced them. The group said only the mathematical community can carry out the assessment the collection requires.
The group described public availability as the beginning, not the end, of the work needed to understand and incorporate the results into mathematics. It also argued that mathematicians must retain the ability to set their own questions and pursue directions not chosen to demonstrate an AI system’s abilities. The group pointed to equitable access to research tools and computing resources as part of that independence.
The test now is mathematical scrutiny
OpenAI says it consulted the advisory group on how to communicate the release and is exploring community-hosted alternatives to its GitHub repository. It also says it plans to fund workshops, conferences and other programs to help people understand major AI-produced results, but has not yet provided details about those efforts.
For now, the public record consists of the manuscripts, supporting artifacts and revision history in the repository. Whether particular results withstand scrutiny, and what mathematical value they hold, is a question the collection leaves to researchers examining the proofs.