OpenAI posted 722 math manuscripts to a public GitHub page on Tuesday. Nearly all of them came from an internal AI model it hasn't released or named. The papers fall into 372 groups of related results. OpenAI says it gave the model about 4,000 problems, and each result took about three hours of computing on average. One result is what OpenAI calls the "quasi-Riemann hypothesis." It's a weaker version of the famous Riemann hypothesis, which OpenAI didn't prove. Lean is software that checks math by computer. By the repo's own count, it has checked the main results in only 162 of the papers. OpenAI says some of the unchecked results "could have issues" and promises to fix them quickly. Last month, an NYU mathematician accused OpenAI of using his team's work to claim a separate result on fluid equations, and OpenAI denied it.

This is a huge leap, and the public should see it

OpenAI and some mathematicians say the results push math forward, and hiding them would make no sense.

This release pushes math forward. OpenAI CEO Sam Altman called it "a new era of discovery." The company wrote in a blog post that it wanted the results "to push the frontier of human knowledge and enable further progress in mathematics." It says future releases will come with better citations and clearer writing.

Even a rival called it historic. Levent Alpöge, a mathematician at Anthropic, posted, "It's obviously the most significant moment in mathematical history." Alpöge worked on the fluid-equation problem with the NYU mathematician who accused OpenAI last month.

OpenAI shouldn't hide the answers. "My view is that this is great for mathematics," University of Toronto math professor Daniel Litt told Fortune. He also told WIRED he sees no reason the company should keep the answers secret. He worries more about how it looks. He fears funders or students could walk away if people think AI has "solved math."

But nobody can check most of it yet

Skeptics say the proofs don't count as science until someone outside OpenAI can test the unreleased model.

Most of these proofs haven't been checked. About 78% of the papers have no computer-checked main proof. Even Lean only confirms the logic holds. It can't say whether a result is new or important. Only 10 of the 372 groups came with a summary of how the model reasoned.

Outsiders can't repeat the work. OpenAI says most results came from "a single prompt handed to a single AI agent." MIT mathematician Andrew Sutherland says people should treat that claim as "unverified" until OpenAI releases the model and others can replicate the results. An independent advisory group hosted at the Institute for Advanced Study told OpenAI to disclose the model, its prompts and the computing time for each result. OpenAI released averages instead. The group says advising OpenAI doesn't mean it endorsed the work.

Nobody knows how the model did it. AI researcher Gary Marcus wrote, "This would never pass peer review." He added that "we don't know what the procedure was" and "we know nothing about the failure rate." He says the system could be a real step toward human-level AI, or "a clever leveraging of Lean and synthetic data" that doesn't carry over anywhere else.

Worse, it hurts how math gets done

Leading mathematicians say math is about people understanding ideas, and releasing hundreds of results at once skips that.

Mathematicians didn't ask for this. The Association for Human Mathematics posted a statement Wednesday on Fields Medalist Terence Tao's blog. It said, "Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power." It urged mathematicians to stop working with OpenAI.

Solving a problem isn't the whole point. In September, 25 Fields Medal winners signed a letter saying AI companies' goals and mathematicians' goals are "severely misaligned." Tao was one of them. They wrote that solving problems is "only a tool and proxy" for understanding. Tao wrote on Mastodon that once they hit their target, the people solving problems now are "AI prompters who have no interest in the broader field itself."

Nobody has settled who gets credit. Tristan Buckmaster, the NYU mathematician behind last month's accusation, told The New York Times he doesn't think OpenAI has done "due diligence at all" on plagiarism. OpenAI has said it can't rule out that data from when he and Alpöge used its products helped improve its models.

Where This Lands

OpenAI and its backers say 722 new results belong in public, and Litt says hiding them would make no sense. Skeptics like Sutherland and Marcus say most of the proofs are unchecked, and nobody outside OpenAI can test the model. The Association for Human Mathematics wants mathematicians to stop working with OpenAI. Tao warns that AI prompters are solving problems without caring about the field around them. Mathematicians now have hundreds of papers to check, and OpenAI says it'll fix any errors they find.

Sources