OpenAI says its models proved local smoothing in every dimension and refuted the Exponential Time Hypothesis
A post dated 26 October describes three results, in harmonic analysis, complexity, and quantum field theory, and three different ways of showing them. The mathematics and complexity writeups and Lean files are public. A physics group led by David Gross is holding its artifacts. Alex Iosevich told Al Jazeera the local-smoothing claim looked historic. He is not named among the shepherds.
Deep dive4 minIn-depth
James Okonkwo4 min read
- OpenAI
- Mathematics
- Complexity
- Physics

Photo: Jeff Chiu / AP, via Al JazeeraSource page
OpenAI said on Monday that its models had produced new results in harmonic analysis, complexity theory, and quantum field theory, and that it is trying three different ways of handing those results to specialists. Al Jazeera wrote the announcement up on Tuesday. The post treats the manuscripts as raw research data. Outside experts, whom OpenAI calls shepherds, advised the release plan. For the mathematics result, the advice was to publish the model’s output and let researchers organize themselves. For the computer-science result, an outside group will explain the work, and the model’s output is public. For the physics result, an outside group will explain it, and that group has decided to hold the artifacts while it writes them up.
Local smoothing, and who has not signed on as a shepherd
The mathematics claim is about the local smoothing conjecture, which Christopher Sogge proposed in 1991. OpenAI’s account: the conjecture says that the most extreme focusing of waves is short-lived, and that it implies the Kakeya conjecture and Stein’s restriction conjecture. The post says Guth, Wang, and Zhang proved local smoothing in two spatial dimensions in 2019, and that Wang and Zahl proved the Kakeya conjecture in three dimensions in 2025. OpenAI says its models produced a proof of local smoothing in every dimension, and therefore also resolve Kakeya and Stein’s restriction in all dimensions, using what it calls a novel auxiliary partial differential equation. It says the model output was obtained on 18 October 2026, formalized in Lean the next day, and then shown to an expert whose recommendation was to release it, the same approach OpenAI says it used on 6 October. A natural-language writeup and a Lean proof are in a public GitHub repository.
The conjecture was formulated by my PhD adviser, Chris Sogge, so it is near and dear to my heart, and the news was a shock at first.
— Alex Iosevich, University of Rochester, to Al Jazeera
Iosevich, a harmonic analyst at Rochester, told Al Jazeera that OpenAI appeared to have cracked the conjecture, and called that a historic development. He said he felt joy and optimism, and also “a twinge of sadness,” and that many theoretical mathematicians he knows feel the same way. He said he expects those feelings to give way to a sense of how much can now be done, and that problems in pure mathematics often came from physical questions that can now be explored quickly. OpenAI’s post does not name him. The lab says this result was released without the standing outside group it assembled for the other two, because the expert it consulted recommended that.
A complexity conjecture, with its authors in the room
The complexity claim is that the Exponential Time Hypothesis is false. OpenAI describes the hypothesis as a longstanding conjecture about the complexity of Boolean satisfiability, widely believed, and used in hundreds of papers to derive conditional hardness results. It says its system produced a natural-language disproof on 16 October and a Lean formalization on 17 October. The refutation, the post says, reopens questions about how much faster certain tasks could become, and “finding such algorithms, and turning theoretical advances into practical improvements, will require significant further research.”
The group OpenAI names to verify and shepherd that result is led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego and Ryan Williams of MIT. The post identifies Impagliazzo and Paturi, with Francis Zane, as the original proposers of the hypothesis. It says OpenAI agreed to two requests from the group: that all raw data be public, and that OpenAI fund a conference on the implications. The natural-language writeup and the Lean proof are in the public repository. Participation rules, which OpenAI says apply to the external collaborations, are that people will not be asked to sign nondisclosure agreements, that OpenAI does not seek academic credit for discoveries produced by its internal models, that the outside group should be credited for the explanations it writes, and that the work is voluntary and unpaid. Members do not speak for OpenAI.
The physics files stay with the physicists
The physics claim is an existence proof, in OpenAI’s words, of certain interacting quantum field theories in four dimensions that contain gauge fields and fermions, including quantum chromodynamics and a unified completion of the Standard Model. The post says the work builds on Balaban’s approach in constructive quantum field theory, and that a subproblem includes a proof of the existence of pure Yang-Mills theory, which is part of the Clay Mathematics Institute’s Millennium Prize problem on quantum Yang-Mills. It does not say the prize problem, as posed, has been closed. David Gross will lead a group that includes Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov. OpenAI’s dates: a natural-language proof for Yang-Mills existence on 9 September 2026, a proof for quantum chromodynamics on 18 September, and a finished Lean formalization of Yang-Mills existence on 16 October. The group, OpenAI says, has decided not to release the model’s artifacts yet. The post prints an address for questions to any member of that effort: YANGMILLS4D@gmail.com.
OpenAI also points to a $5 million contribution to the Simons Foundation, managed and awarded by the foundation, for workshops and programs on results produced by AI. It says choosing the first shepherds “necessarily shapes who gets involved first and creates asymmetry of information.” That sentence is the lab’s own description of the trade. A Lean file that checks is a check of the statements that were formalized. It does not record how the argument was found, and it is not a journal’s decision that the prose claim and the formalization are the same result. Two of the three packages can be read now, on OpenAI’s account. The third is with a named group that has not put the files out.


