Less than a month after it published hundreds of machine-written manuscripts on old mathematics questions, OpenAI is back with something even bolder. On Monday the company announced three new results that it describes as major discoveries: one in pure mathematics, one in theoretical computer science and one in mathematical physics. If even one of them survives expert scrutiny, it will be a moment that researchers talk about for years. If they do not, it will become a lesson in how not to announce science.
The most eye-catching claim comes from computer science. OpenAI says its model has shown that the Exponential Time Hypothesis, usually shortened to ETH, is false. The paper describes a deterministic algorithm that exactly counts the satisfying assignments of fixed-width CNF formulas in subexponential time while using only polynomial memory. Applied to the famous 3-SAT problem, that would refute ETH outright.
To understand why that matters, it helps to know what ETH actually says. Proposed in 1999, the hypothesis states roughly that some hard logic puzzles, such as deciding whether a 3-SAT formula can be satisfied, cannot be solved dramatically faster than by trying an exponential number of possibilities. It is a stronger cousin of the belief that P does not equal NP. Over a quarter of a century, hundreds of results in algorithm design and complexity theory have been built on top of it, often in the form “this problem cannot be solved faster than X, unless ETH is false.” If ETH falls, a large part of that map has to be redrawn.
The mathematics result is just as ambitious. According to OpenAI, its model produced a proof of the local smoothing conjecture for the Euclidean wave equation in every spatial dimension of at least two. The argument reportedly relies on an auxiliary partial differential equation. As consequences, the company says, it also settles the Kakeya maximal and dimension conjectures and Stein’s spherical restriction estimates throughout the interior of the conjectured range. These are central open problems in harmonic analysis. Only a few months ago, NYU mathematician Hong Wang received a Fields Medal for resolving the Kakeya conjecture in three dimensions, which gives a sense of how hard the general case was thought to be.
The third claim, in physics, has not been released in full. OpenAI describes it as a mathematically rigorous existence proof of certain interacting quantum field theories with gauge fields and fermions in four dimensions, including quantum chromodynamics, the theory of the strong force, and what it calls a unified completion of the Standard Model of particle physics. Rigorously constructing such theories in four dimensions is one of the great unsolved problems of mathematical physics, and a version of it is listed among the Clay Millennium Prize problems.
Shepherds, open letters and a new way of releasing results
What is different this time is not only the content but the process. After its October release drew criticism, OpenAI says it is now taking advice from scientists on how to put such results into the world. For the computer science and physics claims, it has recruited outside experts it calls “shepherds” to guide verification and explanation.
The ETH result will be examined by an independent working group led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego, two of the researchers who originally proposed the hypothesis in 1999, together with MIT professor Ryan Williams, one of the best-known names in complexity theory. At the group’s request, the AI-written paper and a formal proof in the Lean proof assistant were published on GitHub immediately, so that anyone can start checking them. A Lean formalization is important: if it compiles and the statement is formalized correctly, the logical steps have been verified by a machine, although experts still need to confirm that the formal statement really means what the paper claims.
The physics work will be shepherded by a group led by Nobel laureate David Gross, including Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. For the mathematics result, OpenAI says the advisers recommended a different path: release the raw model output directly, as in the October 6 release, and let researchers work out their own explanations.
Alongside the papers, the ETH working group published an open letter asking the research community for input. It raises three broad questions. What expectations should govern early access to AI-generated results, disclosure of conflicts of interest and the use of unpublished material in later research? How should credit recognize both a discovery and the hard work of checking, fixing, simplifying, explaining and extending it? And how can students and researchers be supported as methods change, while keeping the understanding and judgment needed to question results and develop new ideas?
Not everyone is convinced by the shepherd model. A panel of mathematics advisers that OpenAI itself assembled, known as AGMAI, said just a day earlier that it recommends AI companies do not adopt the practice of handing results to selected people first. The panel pointed to unequal access, borrowed reputation and biased selection as risks. Instead, it urged AI labs to publish output on scholarly repositories, include Lean-formalized proofs, explain how the problems were chosen and fund the people who do the work of understanding the results. OpenAI acknowledges that its new approach could have its own problems.
That tension goes to the heart of the matter. Traditional science rewards the person who proves a theorem, and peer review is slow precisely because checking is hard. When a machine can generate a long, dense argument in hours, the bottleneck moves from discovery to verification. The scarce resource is no longer clever ideas alone but expert attention. Who gets to look first, who gets credit for finding the errors, and who decides when a result counts as accepted are no longer abstract questions.
There is also good reason for caution. Extraordinary claims, especially ones that overturn widely held beliefs such as ETH, have a long history of collapsing under careful review. Subtle mistakes in definitions, hidden assumptions or gaps between an informal argument and its formal version can take months to surface. Many complexity theorists have long believed ETH to be true, so a refutation would be a genuine shock, and the community will want to test the algorithm on its own terms before rewriting textbooks. The physics claim, with its full details still unpublished, will face even more skepticism until the community can read it.
At the same time, the involvement of researchers of the stature of Impagliazzo, Paturi, Williams and Gross signals that these results are being taken seriously rather than dismissed out of hand. Whatever the final verdict, the coming weeks will be an unusual live experiment in how science absorbs results that were not produced by human hands.
For now, the honest summary is simple: three remarkable claims are on the table, the evidence is public or partly public, and the people best placed to judge them have started work. Whether this goes down as the week AI changed mathematics or as a reminder that proofs still need patient human checking, the way the results are examined may end up mattering as much as the results themselves.





