- OpenAI says its models proved local smoothing for the Euclidean wave equation in every dimension, a result that would imply versions of the Kakeya and Stein restriction conjectures. <sup>[1]</sup>
- The public Lean documentation describes formalized subexponential-time results for fixed-width Boolean formulas, while a group led by researchers associated with the Exponential Time Hypothesis reviews whether the work refutes the hypothesis. <sup>[2]</sup>
- OpenAI says it has an existence proof for four-dimensional theories including QCD and pure Yang–Mills, but the physics outputs have not been released openly. <sup>[3]</sup><sup>[4]</sup>
- Independent tracking of OpenAI’s earlier mathematics release distinguishes machine-checked statements from full validation of the papers’ advertised claims. <sup>[5]</sup><sup>[6]</sup>
OpenAI on October 26 announced three AI-generated results spanning harmonic analysis, computational complexity and quantum field theory, presenting them as raw research that must now be unpacked by mathematicians and physicists. The company says its models produced a proof of local smoothing in every dimension, a disproof of the Exponential Time Hypothesis and an existence proof for several four-dimensional interacting quantum field theories, including quantum chromodynamics. [1]
The release uses different verification paths for the three claims. OpenAI has published natural-language manuscripts and Lean materials for the local-smoothing and ETH work, while it says academic groups will shepherd the physics result. The Institute for Advanced Study’s Advisory Group on Mathematics and Artificial Intelligence says the natural-language and Lean outputs for the Yang–Mills and QCD claims should be made public without delay. [4]
Two mathematical claims have public artifacts #
The harmonic-analysis result concerns Sogge’s local smoothing conjecture, proposed in 1991. OpenAI says the proof uses a new auxiliary partial differential equation and establishes the conjecture in every spatial dimension. The company says local smoothing implies the Kakeya and Stein restriction conjectures, extending results that had previously been proved only in more limited settings. OpenAI points to a 2019 proof of local smoothing in two dimensions and a 2025 proof of the Kakeya conjecture in three dimensions as prior milestones. [1]
OpenAI’s Lean documentation describes a specific formal scope: for every dimension n greater than or equal to 2 and every p from 2 up to, but not including, infinity, it claims a bound for the Euclidean half-wave with the stated Sobolev regularity. The documentation also lists formalized Kakeya maximal estimates and full Hausdorff dimension for Kakeya sets. These are the statements the repository says have been encoded; they are not, by themselves, an independent assessment that the paper’s argument is correct or that its novelty claims hold. [3][6]
OpenAI’s second mathematical claim targets the Exponential Time Hypothesis, a widely used assumption that fixed-width satisfiability problems cannot be solved in subexponential time. The company says its model found a natural-language disproof on October 16 and that the result was formalized in Lean the following day. The public documentation describes a deterministic machine that exactly counts satisfying assignments for fixed-width formulas in subexponential time, with a sharper bound when the number of clauses is linearly bounded by the number of variables. [1][2]
Reviewers have been named, but conclusions have not #
Russell Impagliazzo and Ramamohan Paturi of the University of California, San Diego, and Ryan Williams of MIT will lead the review of the ETH claim. Impagliazzo and Paturi helped formulate the hypothesis, while Williams has worked on related algorithms and lower bounds. OpenAI says the group requested public raw data and that the company will fund a conference on the result’s implications. [1]
That process matters because a Lean file can verify a formal derivation without settling every question surrounding a research claim. Independent commentary on OpenAI’s earlier AI-generated mathematics has emphasized that experts must still check whether the formalized theorem matches the paper’s headline, whether the assumptions are appropriate and whether the work properly credits prior results. [6][7]
An independent evidence-tracking project describes the ETH manuscript as an unverified claim and says that, even if correct, it would not establish P = NP or produce an immediate practical speedup for SAT solvers. OpenAI likewise says converting the theoretical result into useful algorithms will require substantial further research. [5][1]
The physics announcement is less open for now #
OpenAI says its models produced existence proofs for four-dimensional interacting quantum field theories containing gauge fields and fermions, including QCD and a proposed unified completion of the Standard Model. It also says the work includes a proof of pure Yang–Mills existence, one component of the Clay Mathematics Institute’s Yang–Mills Millennium Prize problem. The company says the approach builds on constructive quantum field theory work by David Balaban. [1]
The claim is narrower than solving the full Millennium problem. The Clay challenge asks for a mathematically well-defined quantum Yang–Mills theory and a mass gap; an existence result alone does not settle the mass-gap requirement. Michael Douglas’s 2026 review in Nature Reviews Physics describes those as separate elements of the problem. [8]
David Gross will lead a group including Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov in examining the physics work. OpenAI says it completed a Lean formalization of Yang–Mills existence on October 16, after generating a natural-language proof in September. But OpenAI has not released the full physics outputs in the same way as the mathematical files. The Institute for Advanced Study advisory group says that delay has no scientific justification now that OpenAI says the formalization is complete. [1][4]
OpenAI is building a review system around its releases #
The announcement follows OpenAI’s October 6 publication of 722 mathematical manuscripts covering 372 problem families. An independent evidence project says those totals are counts of published claims, not validated proofs, and that its review of the new ETH claim had not begun as of October 26. The same project says the QCD and Yang–Mills outputs were still awaiting public release. [5]
OpenAI’s broader process follows criticism of its earlier release. A separate independent review of the October 6 repository found that, among formalized families it examined, some Lean statements were narrower than the corresponding paper headlines. That does not establish an error in the three October 26 claims, but it reinforces why the current results should be described as claims under review rather than completed discoveries. [6]
OpenAI also said it is contributing $5 million to a Simons Foundation initiative supporting workshops, conferences and programs for researchers examining AI-generated mathematics. The money may create review capacity, but it does not substitute for independent verification of any individual proof. [9]
Companies mentioned #
Further sources #
The stories that matter, in one email. Free — unsubscribe anytime.