On October 26, 2026, OpenAI announced three results obtained by its internal frontier models in mathematics, computer science, and physics: a proof of Christopher Sogge's 1991 local smoothing hypothesis for all dimensions, a disproof of the Exponential Time Hypothesis (ETH), and rigorous proofs of the existence of QCD and pure Yang–Mills theory in four dimensions. The claims in mathematics and computer science are already supported by machine verification in Lean and the open repository openai/math, and along with the results, the company has for the first time demonstrated a pipeline for publishing AI science: manuscripts are released as "raw research material," undergo formal verification, and are independently reviewed by external expert groups without NDAs.

image
image

What Happened

The announcement covered three areas simultaneously. In mathematics, the proof of the local smoothing hypothesis was obtained on October 18 and formalized in Lean on the 19th, then uploaded to the open repository openai/math; according to OpenAI, the result implies the Kakeya conjecture and Stein's restriction hypothesis. In computer science, a disproof of ETH, the Exponential Time Hypothesis regarding the satisfiability of Boolean formulas, was constructed: the result was obtained on October 16 and formalized on the 17th, with its review assigned to the Impagliazzo–Paturi–Williams group, which includes Russell Impagliazzo and Ramamohan Paturi from UCSD—co-authors of the hypothesis itself. In physics, rigorous proofs of the existence of QCD and pure Yang–Mills theory in four dimensions were claimed: the manuscripts are dated September 9 and 18, the Lean formalization of the existence of Yang–Mills was completed on October 16, and the review is being conducted by a group led by Nobel laureate David Gross. Under the new scheme, all manuscripts are considered "raw research material" and are reviewed by external expert groups without NDAs. The company also added $5 million to the Simons Foundation's $15 million campaign to assist mathematicians in the AI era.

Context

The scale of the claims is easier to assess through the history of the problems. Sogge's hypothesis was formulated in 1991 and describes the behavior of waves during focusing: the strongest focusing should be short-lived. The adjacent Kakeya and Stein restriction hypotheses had previously been closed only in specific dimensions—in the two-dimensional case in the 2019 work by Guth–Wang–Zhang, and in the three-dimensional case in the 2025 work by Wang–Zahl—so the claim of a proof for all dimensions immediately appears extraordinary against the backdrop of progress of one or two steps over years. ETH, the Exponential Time Hypothesis, has served for decades as a foundation for conditional complexity proofs in hundreds of works on graph theory, computational biology, databases, and geometry: conclusions about the computational hardness of problems in these areas were derived from it. The second element of context is the verification tool itself. Lean, an interactive proof checking system where the correctness of each step is confirmed by a computer, shifts the evaluation of scientific claims from the "trust the announcement" mode to the "verify the artifact" mode. Finally, the publication scheme is described as a separate system: independent shepherd groups operate on principles of transparency, OpenAI's refusal of academic credit, and the independence of participants, with the open repository serving as the venue for review with assigned subject-matter experts instead of a journal editorial board.

Why This Matters for the Industry

The industry resolution lies in the experts' verdicts. If the disproof of ETH passes independent review, conditional results relying on the hypothesis will lose their foundation, and the question of how much faster the corresponding tasks can run will shift from conditionally closed to open; practical algorithms from the disproof will not appear automatically, as they will require separate work. If the results hold up, frontier models will be established as a tool for research-level mathematics, and the fall of the 1991 hypothesis within days of obtaining the proof indicates the potential acceleration of the R&D economy. Perhaps more important than the theorems themselves is that the industry now has a candidate for a standard for verifying AI discoveries: if the combination of "raw manuscript plus formal verification plus independent shepherd group" takes hold, machine-verifiable results may emerge as a separate category of scientific publications, and formal-verification-as-eval may become the norm for critical agentic systems. Teams with competencies in Lean and formal methods currently have a window for early positioning: the niche of AI research verification is free until demand forms, and confirmed verdicts will create demand for formal verification tools and theorem-proving agents. From the perspective of immediate deployment, there is no reason to change pipelines: the sources contain no API, pricing, or data on latency and inference costs. However, a transferable pattern is already visible: everywhere a model's output can be automatically verified, an external oracle—tests, compilers, formal schemes—is connected to the agent, creating the same combination of generation and machine verification.

Why This Matters for Users

For the reader, the practical effect right now is reading and verifying with their own eyes, without trusting the announcement. Preprints and Lean proofs are in the open repository github.com/openai/math: for local smoothing, these are the files lean/docs/079.md and paper.pdf; for ETH—lean/docs/378.md. There are no open materials for physics, so this section of the package cannot yet be verified. Next, it is worth following the verdicts of the expert groups and the conference promised by OpenAI on the consequences of the ETH disproof. Loud headlines like "AI solved a Millennium Problem" should be treated with restraint: machine verification has so far been done for mathematics and ETH, while independent review and recognition by the scientific community are still ahead.

What Is Still Unknown / Limitations

The generation methodology has not been disclosed: it does not follow from the sources which specific model and pipeline were used, or what the model's contribution share was in each of the three results. There is also no data on costs, so several loud results are a single case, not a reproducible R&D mode, and the thesis about compressing the cycle from discovery to machine-verifiable proof should only be acknowledged after the results withstand external verification. Independent reviews are just starting: the conclusion of the Impagliazzo–Paturi–Williams group on ETH and the output of the physics group have not yet been issued, and there are also no independent audits of Lean artifacts by the community yet.

Sources

Author

Look at AI, editorial team