OpenAI's AI Disproves 30-Year ETH Hypothesis, Solves Kakeya Conjecture

OpenAI Announces Three Major AI-Generated Scientific Discoveries

On October 26, OpenAI revealed that its internal model produced significant new results across mathematics, computer science, and physics. President Greg Brockman called them "three major scientific discoveries" and noted AI is beginning to help humanity deepen its understanding of the universe. The findings build on the 722 mathematical manuscripts (372 result families, 17 research directions) OpenAI released less than three weeks earlier.

Computer Science: Disproving the Exponential Time Hypothesis (ETH)

Background: 3-SAT and ETH

The Exponential Time Hypothesis, proposed in the late 1990s by Russell Impagliazzo, Ramamohan Paturi, and Francis Zane, asserts that no algorithm can solve all 3-SAT problems in subexponential time. Concretely, the worst-case running time cannot be reduced below 2^{o(n)}, where n is the number of Boolean variables. For nearly 30 years, ETH has served as a foundational assumption; hundreds of papers have derived conditional hardness results for graph algorithms, database queries, and computational biology based on it.

AI's Claimed Counterexample

OpenAI's model constructed a deterministic algorithm for fixed-width CNF Boolean formulas that not only decides satisfiability but also exactly counts the number of satisfying assignments. For instances where the number of clauses is proportional to the number of variables, the algorithm runs in 2^{O(n/√log n)} time (ignoring polynomial factors), which is subexponential. If the proof holds, ETH is falsified. This does not mean all hard problems become easy, but a long-standing theoretical barrier may have been broken.

Review Process

OpenAI invited the original ETH proposers — Impagliazzo (UC San Diego) and Paturi (UC San Diego) — along with Ryan Williams (MIT), a leading researcher in algorithms and complexity lower bounds, to organize an independent, open review. The model generated the natural-language disproof on October 16, completed Lean formalization on October 17, and the result was announced on October 26. Williams and colleagues subsequently announced an open, independent verification effort.

If the proof stands, it means ETH is disproven. This does not imply all hard problems can be solved quickly, but a theoretical boundary thought uncrossable may have been breached.

Mathematics: Local Smoothing Conjecture and Kakeya in All Dimensions

From Kakeya to Local Smoothing

The Kakeya problem asks how much space a needle needs to rotate through all directions. It connects deeply with harmonic analysis and PDEs. In 2025, Hong Wang and Joshua Zahl solved the 3D Kakeya set conjecture. OpenAI's earlier batch (result family 074) claimed proofs of the 3D Kakeya maximal conjecture and 4D Kakeya dimension conjecture, but without Lean formalization at the time.

New Approach: Auxiliary PDE

The new result targets Christopher Sogge's 1991 Local Smoothing Conjecture, which quantifies how wave focusing cannot persist over time — akin to wave crests on water that momentarily peak but quickly disperse. OpenAI's model found a new auxiliary PDE that pushes the proof to all spatial dimensions. According to OpenAI, this result further resolves the Kakeya conjecture and the Stein restriction conjecture in all dimensions.

Timeline and Release

The model obtained the proof on October 18 and completed Lean formalization on October 19. Both natural-language manuscript and formalization are publicly released. Unlike the CS result, OpenAI did not appoint a pre-review panel; researchers are expected to self-organize examination and verification.

Physics: Rigorous Existence Proofs for 4D Quantum Field Theories

QCD and Yang–Mills Existence

Quantum Chromodynamics (QCD) describes the strong interaction between quarks and gluons. While phenomenally successful in calculations and experiments, constructing it as a mathematically rigorous quantum field theory satisfying axiomatic requirements has remained an open problem. OpenAI claims its model, following Tadeusz Balaban's constructive QFT program, established rigorous existence proofs for certain 4D interacting gauge theories with fermions, including QCD and a unified extension accommodating Standard Model gauge fields and fermions. The work also includes a pure Yang–Mills existence proof — part of the Clay Mathematics Institute's Millennium Prize problem "Yang–Mills Existence and Mass Gap" (the mass gap part was not claimed).

Expert Workgroup

2004 Nobel laureate David Gross (asymptotic freedom) will lead a workgroup including Michael Douglas, Arthur Jaffe (co-author of the official Yang–Mills Millennium problem statement), Igor Klebanov, and Nikita Nekrasov. The model generated the Yang–Mills existence proof on September 9 and the QCD-related proof on September 18. Due to complexity and length, OpenAI waited until October 16 (Lean formalization of Yang–Mills existence) before initiating external collaboration. The workgroup will first digest, restructure, and write their own research text before deciding on publication; the raw AI-generated proofs are not yet public.

Three Different Release Strategies for AI-Generated Research

OpenAI frames these outputs as "raw research data," separating proof generation from scientific understanding. The latter requires human researchers to verify arguments, rewrite explanations, connect to prior work, judge contributions, and integrate new tools into existing knowledge frameworks. Accordingly, three distinct mechanisms were adopted:

Mathematics: Direct public release of manuscripts and Lean formalizations; community self-organizes study.

Computer Science: Public release plus invitation to ETH originators and domain experts for independent review.

Physics: Materials handed to an expert team for digestion and rewriting before any public release.

OpenAI stated collaboration principles: voluntary, independent participation; no NDAs; no academic credit claimed for model discoveries; recognition for human explanation, proof curation, and follow-up research. The company pledged $5 million to the Simons Foundation for workshops, conferences, and research projects to support independent review.

When AI can produce large volumes of frontier research in short time, how should the scientific community digest these new conclusions?

The ETH case is symbolic: the very scientists who helped erect a foundational assumption 30 years ago are now invited to examine an AI-generated refutation. For OpenAI, model work ends at proof generation and formalization; for science, the real work — reviewing algorithms, understanding new methods, assessing impact — continues.

References:

Greg Brockman's X post: https://x.com/gdb/status/2114837059462627533

OpenAI blog: Building Scientific Understanding with Academic Communities: https://openai.com/index/building-scientific-understanding-with-academic-communities/