OpenAI dumps 722 AI math preprints to GitHub, pulls three

On October 6, OpenAI uploaded 722 preprints to a GitHub repository, each claiming progress on an open math problem. The papers cover 372 sets of results in geometry, algebra and computer science, produced by an unreleased internal model that was pointed at roughly 4,000 problems, Nature reports. The average compute per result is about three hours of model think time.

Only 162 of the manuscripts have their main results formalized in Lean.

Mathematicians were not asked. A group letter circulating on social media says 'Mathematicians did not ask for this work to be done' and that 'Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power'. Alvaro Lozano-Robledo, a number theorist at the University of Connecticut in Storrs, wrote on X that the simultaneous drop was 'staggering,' asking, 'Why is OpenAI trying to solve hundreds of math problems to be released all at once?' One physicist called the dump a 'slopocalypse'.

A day later OpenAI withdrew three manuscripts after a sign error invalidated the argument in one paper and the construction used by two dependent papers, Retraction Watch reported. Fourteen others were revised with proof repairs. Dan Roberts, a research lead at the company, said OpenAI will 'continue to update the repo with new formalizations and with any errata we notice.'

Alex Townsend, an associate professor of mathematics at Cornell University, said OpenAI picked the wrong order. 'Given the skepticism of AI in the mathematics community, I think that OpenAI should have announced the manuscripts that were Lean verified first,' he said. Andrew Sutherland, a senior research scientist at MIT, called the quick retractions 'the responsible thing to do,' then added that 'it will take a lot more than that to earn back the trust they have lost.'