OpenAI releases 372 math results from its unreleased AI model
OpenAI on Tuesday evening posted 722 papers covering 372 math results made by an AI model it has not released to the public. The company says the work solves or advances major open problems. Mathematicians say checking it will take months, and some want proof of how the model did it.
OpenAI has dropped a pile of new math on the world. On Tuesday evening, the company posted 722 papers grouped into 372 results, all produced by an AI model it has not released to the public. OpenAI says each one solves or makes real progress on a major open question in math or computer science.
The files went up on GitHub at 6 p.m. EDT, Scientific American reported. The claims include a solution to the four-dimensional Kakeya conjecture, faster versions of some key computer algorithms, and progress toward the Riemann hypothesis, one of the hardest problems in math.
Many of the proofs were checked in Lean, a program that tests whether each logical step holds up. That makes those results very likely correct. Whether they hold new ideas, or mostly remix old tricks, will take months to sort out.
OpenAI says the model was given roughly 4,000 problems, according to India Today. The average result used about three hours of ChatGPT Pro thinking time, the company told The Verge. A spokesperson told Scientific American that most results came from a single prompt to a single AI agent, though some may have taken more than one try.
That is where the doubt starts. "We should ask for receipts," said Andrew Sutherland, a mathematician at MIT. He said claims about solving problems in one shot should be treated as unverified until the model is released and others can repeat the work.
Last month, an advisory group of mathematicians formed to guide these releases asked AI companies to share the model, the exact prompts and the computing cost behind each result. OpenAI is sharing only average computing time and some statistics. The spokesperson said the company is not bound by the group's advice.
Not everyone is worried. "To me, it's going to be a good thing for mathematics," said Daniel Litt of the University of Toronto.
The release follows OpenAI's claimed solution last month to the Navier-Stokes problem, a famous question about how fluids flow. That effort used a swarm of 10,000 AI agents and cost millions of dollars in computing, Scientific American reported. CEO Sam Altman wrote on X, "We are entering a new era of discovery now." OpenAI says it will fund workshops and conferences to help experts study the results.
What we know: OpenAI has posted 372 results, many already checked by proof software.
What we don't: How many contain truly new ideas, and when outsiders can test the model.
Where outlets agree Scientific American, The Verge and India Today all report that OpenAI released 372 results from a model the public cannot yet use.
Where they differ OpenAI says most results came from one prompt to one agent, while MIT's Andrew Sutherland says that claim stays unverified until the model is released.