OpenAI has posted 722 mathematics manuscripts to a public GitHub repository. It says an internal model it has not released wrote all of them. The company announced the release on 6 October.
The papers are grouped into 372 “families” of related results, according to Engadget and Unite.AI. A family can hold a main theorem along with companion arguments, consequences or alternative proofs. So 722 does not mean 722 separate open problems solved.
What is in the repository
Unite.AI reports that the repository, called openai/math, is released under an Apache-2.0 license. Besides the PDFs it contains Lean proof formalizations and abridged summaries of the model’s reasoning. OpenAI describes the material as output from model-development evaluations on open research problems.
Xenospectrum’s analysis says the topics include work on zeros linked to the Riemann Hypothesis and a manuscript on the properties of pi. It also says Lean formalization covers 162 of the manuscripts, so most of the collection has no machine-checked proof yet. OpenAI says more formalizations will be added.
The company published ten of the reasoning summaries. It also gave an average compute figure: roughly three hours of ChatGPT Pro compute per result, with about 4,000 problems attempted, according to the AI Weekly roundup and Xenospectrum.
One prompt, one agent
An OpenAI spokesperson told Scientific American, as relayed by Yahoo Tech, that almost everything came from a single prompt given to a single AI agent. Some results may have needed several attempts. That detail is central to the debate, because it suggests little human guidance went into the individual papers.
The headline count has also varied. Tech Insider notes that the repository lists 372 families, while the New York Times reported 377. The site does not settle the gap, and the discrepancy is unexplained.
Mathematicians want checks
None of the papers has gone through conventional peer review. Yahoo Tech’s account says some mathematicians describe the claims as unverified. The OpenAI release points to a new advisory group, which Engadget says confirmed some of the results. That group is the Advisory Group on Mathematics and AI, and Tech Insider says it issued responsible-release recommendations on 29 September after more than 600 replies.
Xenospectrum compares the release with those guidelines. The guidelines call for storage in a repository the AI lab does not control, persistent citation identifiers, and a full account of how each result was generated. They also call for proofs to be formalized wherever possible. OpenAI’s repository includes citation information and a revision policy that keeps earlier versions. It is hosted on OpenAI’s own GitHub, though a community-run alternative is under consideration, per that analysis.
The release follows a dispute in September over an OpenAI claim about the Navier-Stokes equations. Engadget says the new batch is likely to feed that controversy. A summary from Splitfeed, an aggregator, says 25 Fields Medal winners have warned that mass-producing results could harm the field, and that a mathematician told Wired of a “perception of mobster behavior” among leading AI labs. Those claims come through secondary summaries, and the underlying statements were not reviewed here.
What is still open
The model is not available, and Tech Insider says OpenAI has given no timeline for releasing it. Outsiders can read the papers and run the Lean checks, but they cannot test the system that produced them.
Whether the individual results hold up will depend on mathematicians reading them, and on how many of the 722 manuscripts end up with complete formal proofs.