OpenAI says its internal models have produced three major research results at once: a refutation of the Exponential Time Hypothesis (ETH), a cornerstone assumption of computational complexity theory; a proof of the local smoothing conjecture in all dimensions; and rigorous existence proofs for certain four-dimensional quantum field theories. The company announced the work on October 26 and published manuscripts and Lean formalizations for the first two. Independent groups of leading researchers are now reviewing the claims, and none has yet issued a verdict.
The ETH Claim
The Exponential Time Hypothesis, formulated in 1999 by Russell Impagliazzo and Ramamohan Paturi, roughly states that the 3-SAT problem cannot be solved in subexponential time. Hundreds of papers in algorithms and complexity theory rely on it as a working assumption to prove that faster algorithms are unlikely.
OpenAI's preprint, titled "The Exponential Time Hypothesis is False," claims a deterministic algorithm that exactly counts the satisfying assignments of k-CNF formulas, for any fixed k, in subexponential time and polynomial space. Applied to 3-SAT, that would refute ETH. According to coverage by Unite.AI, the natural-language disproof was obtained on October 16 and formalised in Lean the following day, with a manuscript dated October 25 giving sharper bounds for instances with a given number of clauses.
Two caveats matter. First, OpenAI says the result does not resolve P versus NP and does not immediately translate into practical speedups; turning it into faster real-world solvers would need significant further research. Second, a Lean formalisation checks that a proof follows from its stated definitions, but experts still need to confirm that those definitions capture the intended mathematical statement.
Who Is Checking the Work
OpenAI used a different release approach for each result:
- Computer science: The ETH paper and Lean proof were posted publicly at the request of an independent group led by Impagliazzo and Paturi of UC San Diego and Ryan Williams of MIT. ForkLog reports that the group stressed its participation is not an endorsement and that it plans an open verification with the wider community.
- Mathematics: The local smoothing manuscript was released directly with its Lean formalisation, without a dedicated reviewer group.
- Physics: Materials went first to a group led by Nobel laureate David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, which will decide how to publish.
Unite.AI reports that OpenAI agreed to make the raw data public and to fund a conference on the implications of the ETH result. OpenAI says participation by outside scientists is voluntary and unpaid, without non-disclosure agreements, while acknowledging that choosing the initial reviewers creates an information asymmetry.
Local Smoothing, Kakeya and Quantum Field Theory
The mathematics claim is arguably as ambitious. OpenAI's repository says a new manuscript proves Sogge's local smoothing conjecture for wave equations in dimensions two and higher, and that it also yields the Kakeya maximal and dimension conjectures and Stein's restriction estimates within the conjectured range. Local smoothing had previously been proven only in two dimensions, and Kakeya only in three, according to Unite.AI. The model output was reported on October 18 and formalised on October 19.
On the physics side, OpenAI says its models built existence proofs for interacting four-dimensional quantum field theories with gauge fields and fermions, including pure Yang-Mills theory and quantum chromodynamics. ForkLog notes that OpenAI is not claiming a full solution to the Clay Millennium problem, which also requires proving a mass gap. The Yang-Mills existence proof was formalised in Lean on October 16, but the QCD extension has not been reported as formalised.
A Track Record Under Scrutiny
The announcement follows a rocky start to OpenAI's research-publishing push. OpenAI's own repository history shows that on October 7 it withdrew three manuscripts after a sign error invalidated a key argument in a paper on Weil classes, which affected two dependent papers. It also repaired proofs or clarified statements in 14 other manuscripts. At that point the repository reported that 300 of 719 top-line results, about 42%, had been formalised.
Researchers criticised the earlier releases for limited pre-review and unclear authorship, ForkLog reports. The new staged approach, with named expert groups shepherding each field, appears designed to address that criticism.
Why It Matters
If the ETH refutation holds up, it would be one of the most consequential results in theoretical computer science in decades, forcing researchers to revisit a large body of conditional lower bounds. It would also mark a turning point for AI-driven research: a machine-generated proof overturning a widely believed conjecture, backed by a formal proof that anyone can check.
The bigger story may be the method. Pairing model-generated proofs with Lean formalisation and independent expert review offers a template for how AI labs can publish claims that are too large to accept on trust. For now, the right posture is caution. The verification groups have only just begun their work, and the field will wait for their conclusions before rewriting textbooks.
