Skip to content

OpenAI

OpenAI hands $5M to the Simons Foundation to vet AI math

3 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: independent.co.uk

OpenAI is putting $5 million into a new Simons Foundation initiative that will pay mathematicians to examine, understand and build on results produced by AI. The company announced it on October 15 in a post on X, and said the work covers the results it published itself nine days earlier.

OpenAI is not running the program. The Simons Foundation manages the money and decides who gets it, and the money goes to workshops, conferences and special programs. Researchers the foundation supports also get access to OpenAI’s models and computing resources. OpenAI has not said how the $5 million splits across those programs or when the first grants go out.

The results in question came out on October 6. They were produced by an unreleased internal OpenAI model, and the company’s GitHub repository lays out the scale. The current catalogue holds 719 manuscripts in 372 families. Most of them came out of one fixed procedure: the model was posed approximately 4,000 problems, and each result used three hours of ChatGPT Pro thinking compute on average. A few results were handled differently, among them a zero-free region for the Riemann zeta function and a proof of the Hodge Conjecture for CM abelian varieties. The write-up of the Re(s) > 11/12 zero-free region was edited by humans for readability. Abridged summaries of the model’s reasoning were released for 10 result families. TechCrunch reported that the same model has resolved more than 100 additional open problems across most areas of mathematics.

That is the material the Simons money is meant to help people check, and the repository says only about 42% of the top-line results have been formalized in Lean, a language whose checker confirms every step of a proof by machine. OpenAI says some of the unformalized results could have issues. One already did: a single sign error led OpenAI to withdraw three papers on October 7, a day after the release, because two of them relied on the same construction. Reporting by Yahoo Tech describes a verification bottleneck, where humans may struggle to check research as quickly as AI can produce it.

The mathematicians’ complaints go further than error rates. Per TechSpot, the objection is that work done on proprietary models puts results ahead of understanding and leaves researchers unable to examine the systems behind the proofs. Some of them question whether they can build on the findings without access to the model. The advisory effort around all this followed an open letter signed by more than two dozen Fields Medal winners, titled “A Severe Misalignment of AI in Mathematics.” Giving Simons-supported researchers access to OpenAI’s models looks like an answer to the access complaint in particular, though the post speaks of “our models” in general and the published results came from a model OpenAI has not released.

OpenAI has also formed a mathematics advisory group, which it describes as independent mathematicians who act as a bridge to the mathematical community and the wider public. For the October 6 release, it said it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study.

That consultation had limits. TechSpot reported that the group gathered feedback from hundreds of researchers and issued publication guidelines, and that OpenAI followed most of them but not all. One of the recommendations asked developers to stop testing advanced mathematical problems on proprietary models. OpenAI turned that one down and defended continuing to evaluate its internal frontier models.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.