The people working on OpenAI’s mathematics program have begun to speak in a way they rarely do, and what they are saying would have seemed improbable even a year ago. According to people familiar with the research, the company is now close to solving another of the seven Millennium Prize Problems, the Hodge conjecture, after having made progress on the Navier-Stokes existence and smoothness problem before it.
The employees closest to the work expect the Hodge conjecture to be resolved soon, the people said. That would make OpenAI the author of a second breakthrough on a list of problems that has yielded only one full solution since the Clay Mathematics Institute posted the million-dollar questions in 2000.
The Hodge conjecture sits deep inside algebraic geometry, the study of shapes defined by polynomial equations. It asks whether certain geometric features of those shapes can always be built from simpler algebraic components, a question that connects the smooth geometry of a surface to the equations that describe it. The conjecture was posed by the Scottish mathematician W. V. D. Hodge in the middle of the twentieth century, and it has resisted every assault since.
OpenAI’s approach differs from the century of human effort that preceded it. The company has trained its models to reason through mathematics at length, and the people familiar with the work say the systems have been exploring the conjecture’s structure in ways that suggest a proof is within reach. The Navier-Stokes problem, which concerns whether solutions to the equations governing fluid flow always behave smoothly, was the earlier target.
A proof of either problem would carry consequences far beyond the prize money. The Navier-Stokes equations underpin everything from weather forecasting to aircraft design, and a resolution would settle a question physicists have lived with for generations. The Hodge conjecture is more abstract, but its resolution would reorganize a corner of geometry that mathematicians have been building around for decades.
The company is in no hurry to publish, the people said. OpenAI is studying how to release the work in collaboration with the mathematics community, a process meant to avoid the public-relations crisis that has attended some of its earlier research announcements. The memory of those episodes is fresh enough inside the company that the timing and manner of any announcement are being treated as carefully as the mathematics itself.
That caution reflects a deeper unease in the field. Mathematicians have watched the AI labs move into their discipline with a mix of fascination and alarm, aware that a machine proof of a Millennium Problem would change the profession’s sense of itself. A result generated by a model, rather than discovered by a human, raises questions about authorship, verification, and what it means to understand a theorem at all.
OpenAI has reason to manage that conversation deliberately. Earlier claims about its models’ capabilities drew criticism that the company had overreached or failed to share enough for others to check the work, and a claim as significant as a Millennium Prize solution would be scrutinized at a level no press release could survive if it fell short.
The mathematics program is part of a broader push at OpenAI to apply its models to hard, verifiable problems rather than to the open-ended text generation that made the company famous. Mathematics is attractive because a proof is either right or wrong, and the correctness can be checked by machine and by human referees alike, giving the company a claim to rigor that chat output cannot offer.
The Clay Institute has said it will award its prizes only to work that survives peer review and the passage of time, a standard designed to prevent premature claims. That means even a solved conjecture would not become official immediately, and the gap between a company announcement and a Clay confirmation could stretch for years.
The seven Millennium Prize Problems were chosen in 2000 as the field’s most stubborn open questions, each carrying a $1 million prize from the Clay Mathematics Institute. In the quarter century since, only one has been solved: the Poincaré conjecture, proved by the Russian mathematician Grigori Perelman in 2003, who declined both the prize and the Fields Medal. The others, including the Hodge conjecture and the Navier-Stokes problem, have withstood every human attempt.
The Navier-Stokes problem, which the people familiar with OpenAI’s work say the company has already approached, asks whether the equations describing fluid motion always produce smooth, well-behaved solutions. A negative answer would mean the equations can break down in ways mathematicians cannot yet predict, with consequences for physics that would reach far beyond the prize.
For now, the work remains internal, and OpenAI has said nothing publicly. What the people familiar with the research describe is a company confident enough in its results to begin planning how to tell the world, and cautious enough to know that the telling matters as much as the theorem.
