OpenAI Releases 722 Math Manuscripts, Single-Prompt Problem Solving Sparks Verification Controversy

722 manuscripts, 372 hard math problems, with each result averaging roughly three hours of ChatGPT Pro usage—that's the scale of what OpenAI has just disclosed. But what's really intriguing is the method: according to what OpenAI told Scientific American, nearly every single paper was produced using one prompt handed to one AI agent.

All of the manuscripts have been uploaded to GitHub, verified by OpenAI's newly formed advisory group, the Advisory Group on Mathematics and Artificial Intelligence (AGMAI). Last month, OpenAI had already signaled its intentions, claiming its model "solved more than 100 long-standing open problems spanning most areas of mathematics." This release essentially lays out the concrete list. The company stated that the results were produced by a ChatGPT pilot model not yet released to the public.

Among the headline results OpenAI claims are a solution to the four-dimensional Kakeya conjecture, improvements to several key computer algorithms, and progress on the Riemann hypothesis—any one of which, if confirmed, would be a major event in the math world. In its announcement, the company wrote that it is starting with a GitHub repository release, complete with mechanisms for paper revisions and citations, "while still seeking other community-hosted options that meet the advisory group's guidelines," and promised that future releases will be more complete in terms of citation standards, mathematical exposition, and presentation.

But OpenAI only followed through on half of AGMAI's recommended transparency. The advisory group had originally recommended disclosing the compute time and prompts used for each individual problem, but OpenAI did not do this, instead only providing an overview of the overall reasoning process, compute estimates, and the number of problems attempted. In other words, the public can see the list of results, but cannot see the specific path that would let anyone verify how those results were actually produced.

This is exactly where the controversy lies. Prior to this, OpenAI's claims related to Navier-Stokes had already stirred up a wave of backlash in the math community, leaving many scholars wary of "one-shot AI problem solving" claims. MIT mathematician Andrew Sutherland told Scientific American: "Unless they release the model so results can be reproduced, any claim of a single agent solving a problem in one shot should be considered unverified."

The batch of manuscripts OpenAI has released still needs to be reviewed and verified by mathematicians one by one, and the real impact won't be clear until that review process concludes. As for whether the model itself will ever be made public, or when the specific prompts will be filled in, OpenAI has not given a timeline.