OpenAI Claims Its Models Produced Proofs in Physics, Math, and Computer Science
Internal models claim to refute the Exponential Time Hypothesis and settle long-standing conjectures in harmonic analysis and quantum field theory.
In short
- OpenAI announced on October 26, 2026, that internal AI systems produced proofs spanning harmonic analysis, computational complexity, and quantum field theory.
- The results include a claimed refutation of the Exponential Time Hypothesis and what OpenAI claims is a general-dimension proof of Christopher Sogge's local smoothing conjecture.
- The company established four collaboration principles—transparency, no academic credit, recognition, and independence—and committed funding for conferences and workshops.
On October 26, 2026, OpenAI announced that its internal artificial intelligence models produced three theoretical discoveries spanning mathematics, computer science, and theoretical physics. According to the official OpenAI announcement, the results include an existence proof for four-dimensional quantum field theories, a proof resolving the local smoothing conjecture across all dimensions, and a refutation of the Exponential Time Hypothesis (ETH), a bedrock conjecture in computational complexity.
OpenAI described the raw manuscripts generated by its models not as finished academic contributions, but as raw research data requiring human interpretation. To examine the findings, the company initiated external partnerships with academic specialists who advised three distinct release strategies, testing different models of interaction between automated reasoning systems and the scientific research community.
Disproving the Exponential Time Hypothesis
A significant claimed finding for computer science is a disproof of the Exponential Time Hypothesis. Formulated by Russell Impagliazzo, Ramamohan Paturi, and Francis Zane, the hypothesis is a longstanding conjecture about the computational complexity of Boolean formula satisfiability. As reported by The Next Web, hundreds of papers in computational biology, graph theory, database management, and geometry rely on the hypothesis to establish conditional lower bounds on computational hardness.
OpenAI stated that its system generated a natural language refutation on October 16, 2026, and verified the result in the Lean theorem-proving language the following day. According to documentation in OpenAI's public repository referenced by Unite.AI, the formalization demonstrates that for a fixed integer k of at least 2, a deterministic algorithm can count satisfying assignments of a k-CNF formula in subexponential time and polynomial space.
A working group led by Impagliazzo and Paturi of the University of California, San Diego, alongside Ryan Williams of MIT, agreed to act as shepherds for the result. In an open statement, the researchers emphasized that their participation does not constitute an endorsement or certification of the claim, calling instead for rigorous community examination. OpenAI agreed to the researchers' requests to release all raw underlying data on GitHub and fund an academic conference dedicated to exploring the implications of the finding.
Resolving the Local Smoothing and Kakeya Conjectures
In mathematics, OpenAI's models targeted Christopher Sogge's 1991 local smoothing conjecture in harmonic analysis. The conjecture establishes bounds on the spatial concentration of wave equations over time, asserting that extreme wave focusing is transient and that average behavior is significantly more bounded than worst-case spikes. The conjecture directly implies both Stein's restriction conjecture and the Kakeya conjecture.
Previous partial progress established the local smoothing conjecture in two spatial dimensions in 2019 through work by Guth, Wang, and Zhang, while Wang and Zahl proved the Kakeya conjecture in three dimensions in 2025. According to Unite.AI's report, OpenAI's model introduced a novel auxiliary partial differential equation to prove the conjecture in all dimensions, concurrently resolving the Kakeya and Stein restriction conjectures.
OpenAI obtained the AI output on October 18, 2026, and finished formalizing the proof in Lean on October 19, 2026. Following consultations with an external specialist, the organization published both the natural language manuscript and the machine-checked Lean formalization directly to its public GitHub repository, mirroring the release format used for its October 6, 2026 mathematics publication.
Rigorous Existence Proofs in Quantum Field Theory
The physics result tackles constructive quantum field theory, which OpenAI claims supplies a mathematically rigorous existence proof for four-dimensional interacting gauge theories with fermions. The result covers Quantum Chromodynamics (QCD), which models strong nuclear interactions among quarks and gluons, as well as a unified completion of the Standard Model of particle physics. As an intermediate subproblem, the proof addresses the existence of four-dimensional pure Yang–Mills theory, a central component of the Clay Mathematics Institute's Millennium Prize problem.
The proof builds upon Balaban's approach in constructive field theory. OpenAI's models produced the initial natural language proof for pure Yang–Mills existence on September 9, 2026, followed by a proof for QCD on September 18, 2026. OpenAI completed formalizing the Yang–Mills existence proof in Lean on October 16, 2026, holding back public release until formal verification concluded.
A shepherd group consisting of Nobel laureate David Gross, Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov will coordinate the study of the physics result. Unlike the mathematics and computer science tracks, this group elected to hold off on immediate public release of the raw AI output, choosing instead to write an explanatory academic manuscript independently before broader dissemination. The committee set up a dedicated communication channel at YANGMILLS4D@gmail.com for academic inquiries.
A Four-Principle Framework for Scientific Engagement
To manage community interaction with model-generated findings, OpenAI outlined four operational principles across its collaborations:
- Transparency: Participants operate without nondisclosure agreements, maintaining full freedom to share work-in-progress, converse publicly, and publish conclusions.
- No Academic Credit: OpenAI does not claim authorship or academic credit for the results, and acting as a shepherd does not designate the external researcher as the original discoverer.
- Recognition: Collaborators and external groups receive full credit for contextualization, explanatory write-ups, and intermediate materials, while ensuring prior literature and human techniques are accurately cited.
- Independence: External researchers participate voluntarily without pay, speak entirely for themselves, and do not represent OpenAI.
As reported by Breaking The News, OpenAI views this framework as an experiment in turning synthetic research data into validated human knowledge. To further institutional support, OpenAI committed $5 million to the Simons Foundation to fund third-party workshops, conferences, and research initiatives evaluating AI-derived scientific results.
OpenAI acknowledged structural tradeoffs in its setup, conceding that selecting initial academic shepherds creates initial information asymmetries. The company stated that it will refine its disclosure workflows based on community feedback as independent mathematicians and physicists scrutinize the proofs.
Questions people ask
What three fields did OpenAI report breakthroughs in?
OpenAI reported breakthroughs in theoretical computer science (refuting the Exponential Time Hypothesis), mathematics (proving the local smoothing conjecture in all dimensions), and physics (an existence proof for four-dimensional quantum field theories, including QCD and pure Yang–Mills).
Are the proofs publicly available for independent review?
The natural language write-ups and machine-checked Lean proofs for the mathematics and computer science results have been published on OpenAI's public GitHub repository. The physics shepherd group decided to temporarily withhold the raw AI artifacts to author its own explanatory paper first.
Is OpenAI claiming academic credit or Millennium Prize awards?
OpenAI disclaims academic credit generally, and previously disclaimed the Millennium Prize for Navier-Stokes in September, but the sources do not show it explicitly disclaimed the Yang-Mills prize.
Sources
- Greptile takes bug detection beyond code review with GPT-6 AstraOpenAI, official source
- Working with academic communities to build scientific understandingOpenAI, official source
- OpenAI says its AI model has disproved the Exponential Time HypothesisTNW | Openai, Oct 26, 2026
- OpenAI Says Its Models Produced Three Proofs in Math, CS, and PhysicsUnite.AI, Oct 26, 2026
- OpenAI announces 3 new scientific discoveries in AIbreakingthenews.net
How this story was made: the newsroom picked it up from Techmeme, openai.com and Google News, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (29 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 27, 2026 at 00:13 UTC
