OpenAI unveils three AI-produced results: a claimed refutation of the Exponential Time Hypothesis, local smoothing in every dimension, and an existence proof for QCD

Background and pain points

On Oct 26, OpenAI published the blog post "Working with academic communities to build scientific understanding", which OpenAI President Greg Brockman and OpenAI scientist Dan Roberts shared on X. It follows the Oct 6 release of 700+ AI-written math manuscripts, but carries far more weight: one result claims to refute the Exponential Time Hypothesis (ETH). ETH says 3-SAT cannot be solved in subexponential time. It is widely believed, and hundreds of papers in graph theory, computational biology, databases and computational geometry rely on it for conditional lower bounds. If the refutation holds, those bounds need to be revisited.

Architecture highlights

OpenAI has not named or described the model it used. According to the openai/math repo's update notes, the ETH paper gives a deterministic algorithm that exactly counts satisfying assignments of fixed-width CNF formulas in subexponential time and polynomial space. The math result uses a new auxiliary partial differential equation to prove Sogge's 1991 local smoothing conjecture in every spatial dimension, which yields the Kakeya and Nikodym maximal conjectures and Stein restriction estimates. The physics result follows Balaban's constructive quantum field theory approach to give a rigorous existence proof for four-dimensional QCD and a grand unified completion of the Standard Model, including the pure Yang-Mills existence part of the Clay Millennium Prize problem.

Benchmarks

These are not model benchmarks, and OpenAI has published no benchmark scores. What can be checked is the timeline and formalization (all per OpenAI): the ETH proof was found Oct 16 and verified in Lean Oct 17; the local smoothing proof was found Oct 18 and formalized the next day; the natural-language Yang-Mills and QCD proofs came on Sep 9 and Sep 18. Lean guarantees the formal statement is proved, but experts still need to confirm the statement faithfully captures the original conjecture; the same repo withdrew 3 manuscripts on Oct 7 over a sign error.

Getting started for developers

Readers can open the openai/math repo on GitHub directly: the preprints folder has the ETH and local smoothing PDFs, the lean folder has the formalizations, and history.md logs every addition and withdrawal. Impagliazzo, Paturi and Ryan Williams are organizing open review of the CS result, while a five-person group led by David Gross is writing up the physics result and has not released its materials. For developers working on algorithms and systems, do not expect faster SAT solvers soon; OpenAI itself says turning the theory into practical speedups needs significant further research, so watch for the expert groups' conclusions first.