On October 6, OpenAI posted hundreds of mathematical manuscripts at once. An unreleased internal model had produced all of them. They went into a public GitHub repository along with computer-checkable versions of many of the proofs and short summaries of the model's reasoning for ten of them. In an episode of Moonshots with Peter Diamandis recorded October 13, computer scientist Alexander Wissner-Gross, founder of Reified, said the release marked the end of an era. "Mathematicians as artisans. It's over. It's cooked," he said.
Not everyone saw a triumph. A group of mathematicians asked colleagues to stop working with OpenAI, and within a day the company had withdrawn three of the manuscripts. The hosts mostly argued about something bigger: whether math was simply the first field to go.
What OpenAI put out
Host Peter Diamandis, founder of XPRIZE and Singularity University, opened the segment with a figure: 722 "mathematic solutions" in a GitHub repository, from a model he said to "call GPT-7 for the moment." Wissner-Gross later called it "Bell." He said the model was reportedly expected to become generally available by the end of the year. Diamandis also relayed a remark by Mark Chen, OpenAI's chief research officer, describing "a decade of mathematical progress in a week." He added that Chen's next target is the life sciences.
OpenAI's announcement describes the release as results from an unreleased internal frontier model. It includes manuscripts, many Lean formalizations and information about the problems the model attempted. Lean is a system in which a proof is written out formally enough for a computer to check every step. The company said the average result took about three hours of ChatGPT Pro-level thinking. It also said it would fund workshops and conferences to help researchers understand major AI-produced results, and would improve citations and mathematical writing in future releases.
What the repository actually counts
The repository does its own accounting, and it is more detailed than the headline number. It separates individual manuscripts from "families" of results. One family can contain a main result plus supporting arguments, consequences or alternative proofs. The revised catalogue lists 719 manuscripts in 372 families. The episode's title and Diamandis both used 722.
According to the repository, the model was given about 4,000 problems. Its outputs were then grouped and filtered for significance. Most results came from the same procedure, averaging three hours of thinking compute per result. The repository names two exceptions: work on a zero-free region of the zeta function, and work on the Hodge conjecture for one special class of objects, CM abelian varieties. OpenAI calls verification incomplete and warns that results without a formal proof may contain problems.
That warning soon came into play. The release history dated October 7 shows OpenAI withdrew three manuscripts. A sign error broke an argument about Weil classes, and two papers that depended on it, involving Kuga–Satake correspondences and products of K3 surfaces, fell with it. The log also records repairs or clarifications to 14 manuscripts and citation updates to 13 more. Additional Lean formalizations raised the count of formally checked results to 300 of the 719, about 42%. The withdrawn papers stay accessible with notices explaining their gaps.
The hosts' numbers
Wissner-Gross described the effort as surprisingly modest. In his account, it took two weeks, "a single model, like a single agent" given at most three hours per problem, "one-shotting" each problem rather than sending a swarm at it. He thought "something like 15 to 20 percent of the problems that were posed were ultimately solved." He contrasted this with his recollection of OpenAI's earlier Navier–Stokes work, which he remembered as "thousands of agents and 88 hours and millions of dollars." He guessed the math effort cost "low tens of millions maximum, maybe a lot less."
The repository partly matches that picture. Three hours is its average, not a ceiling, and it lists exceptions. Compared with roughly 4,000 problems, 719 manuscripts works out to about 18%, close to his estimate. The 372 families work out to about 9%. Neither is exactly a count of solved problems, because one problem can yield several manuscripts. The repository does not support or address the single-agent setup, the two-week timeline or the cost estimate.
Wissner-Gross said at the start of the episode that 81% of "the most important math results over the past three years" were published in a single day, and later repeated it as true "by one measure." Neither he nor Diamandis said what the measure was.
Wissner-Gross also suggested OpenAI "wasn't even breaking a sweat." If it had wanted to take on the Riemann hypothesis or the Hodge conjecture, he said, it "probably could have spent six hours or 10 hours" to solve them. That is his speculation. The release does not claim either result.
Diamandis raised two rumors. One was that more solutions were coming. The other was talk that some solutions may have leaked from mathematicians' earlier conversations with ChatGPT. Wissner-Gross said he would "intentionally avoid commenting on that for the time being." Nothing in OpenAI's materials addresses the leak claim.
What the mathematicians objected to
The objection came from the Association for Human Mathematics. Its Communications Working Group issued a statement on October 7, republished as a guest post on Terence Tao's blog. The association said OpenAI had tested advanced problems internally despite the stated position of an advisory group. It also objected to the company publishing hundreds of mathematical documents together. The statement's argument rests on scientific norms, public trust and human understanding, rather than treating the number of solutions as the only measure of progress. It urged mathematicians to stop working with OpenAI.
On the show, Diamandis read a line he attributed to the group: releasing over 700 files at once "is not a demonstration of scholarship, but a demonstration of power." He also relayed comments from MIT's Andrew Sutherland, who said many mathematicians are "unhappy with how OpenAI has behaved" and that "it will take a lot more to earn back the trust that they've lost."
Wissner-Gross was unsympathetic. He called the association's name "pathetic" and its members "bellyachers in the field who were counting on personal glory and artisanry being rewarded." He told unhappy mathematicians to "get a life." He said Tao and others had taken a position that the release was bad for the practice of math. Then he steelmanned that position in his own words: the AI only solves problems and doesn't pose new ones or new conjectures, so it is "stealing the seed corn of the field." The association's statement focuses on process, disclosure and trust, not on that argument.
"Math is cooked," and everything else next
Wissner-Gross's larger claim builds on Solve Everything, a blueprint he co-wrote with Diamandis. It proposes turning whole domains into problems that computation, data and repeatable testing can work through, which they call bulk solution. It is a proposal, framed as an organizational and engineering program, including infrastructure to verify results and payment for outcomes. In his view, the October release was the first time an entire field had been bulk-solved. "Math is cooked. And every other field is about to be cooked," he said. Math went first, he argued, "because math is easiest to verify."
Diamandis asked the question many listeners would: are these just obscure problems? Wissner-Gross said math is the foundation of science and engineering. "If we can solve math, which we now arguably have, we can solve everything else." He listed physics, chemistry, biology and medicine, and added that "the humanities even depend on math, even though they're in denial." He called the moment one of the singularity's inflection points and said history would treat it as "a historic watershed." He meant the point at which an entire field "was bulk solved."
Why Dave Blundin thinks mathematicians are fine
Diamandis asked Dave Blundin, founder of Link Ventures, whether MIT's mathematics department was screaming. Blundin said "it's actually the opposite." In his telling, mathematicians are "in great shape because they're so cooked" that they are becoming AI natives fast. The people he hears are slower to react are in biotech and deep in the physics labs, "because they don't feel cooked yet." He described the prevailing attitude at MIT: "these proofs are ugly and not artisanal and not subtle, but they're right." That leaves a hybrid period in which humans are still useful alongside AI. He said that window "may only last a couple of years."
Blundin explained why solved math would matter outside math with a memory from early personal computers. When you opened an Apple II and wanted to draw a circle, he said, there were no toolkits. You cracked open a math book and spent "three or four hours" working out the math behind a single circle. Once math is solved, he argued, "it's a simple lookup of the answers." AI could then move quickly through engineering problems that are "all math under the covers," like 3D design, rocket design, gears that mesh and robots that move correctly.
Salim Ismail: the research team is over
Salim Ismail, founder of Open ExO, began with a joke. He proposed a Monty Python "dead parrot" compilation of Wissner-Gross declaring math "cooked" and "pushing up the daisies." Wissner-Gross added, "It is an X field." Then Ismail said he felt personally vindicated. As a theoretical physics student, he recalled, professors took him aside and warned that he would never make it as a physicist without getting his foundational mathematics right. He never did, preferring philosophy courses on the side.
His broader point was about how research is organized. A tenured professor used to lead a small team of postdocs and students, he said, but "now one researcher can direct hundreds of agents and totally solve it." His conclusion: "Every PhD, advanced PhD program in the world is toast right now." Wissner-Gross replied that it was "even worse," or more impressive, than that. This release, he said, didn't even use swarms of agents.
The lesson Blundin draws for other fields
Blundin took a practical lesson from that point. In his view, there is much to gain from "the looping and prompt management that they figured out with Navier, Stokes, and then were able to apply after that." First attempts are brute force and full of mistakes, he said. Then teams get much better at how they reprompt, loop and use the context window, meaning the material a model can hold in view at once. He called it "a good roadmap for the material scientists, the physicists, the biologists, everybody, even the movie makers." With really good looping and prompting, he said, a filmmaker's efficiency in making an entertaining movie "can go up a factor of 10 to 100."
Listen to the full discussion on the episode page.