OpenAI announced on October 26, 2026 three claimed discoveries—in mathematics, computer science, and physics—and a release process guided by external “shepherds.” The company frames the AI manuscripts as raw research data that still need scientific understanding. External experts advised three different strategies: release the math AI output for the community to explain; form a CS working group while publishing the AI output immediately; and form a physics group that defers release of AI artifacts. OpenAI lists four collaboration principles: transparency (no NDAs), no academic credit sought by OpenAI for the model discoveries, recognition for the understanding groups, and unpaid independent participation.
Key Facts
| Domain | Claim (OpenAI) | Public materials (as of announcement) | Shepherd / release note |
|---|---|---|---|
| Mathematics | Proof of Sogge's local smoothing conjecture in every dimension; implies Kakeya and Stein restriction in all dimensions | Natural-language writeup + Lean proof on OpenAI GitHub | Expert advised direct release (same approach as Oct 6, 2026); AI output dated Oct 18, Lean next day |
| Computer science | Exponential Time Hypothesis (ETH) is false; Lean formalization claimed | Natural-language writeup + Lean proof on GitHub | Group led by Russell Impagliazzo, Ramamohan Paturi (UCSD), Ryan Williams (MIT); materials public at group's request |
| Physics | Rigorous existence for certain 4D gauge+fermion QFTs including QCD and a unified Standard Model completion; subproblem includes pure Yang–Mills existence (Millennium Prize problem component) | AI artifacts not released yet | David Gross leads group with Michael Douglas, Arthur Jaffe, Igor Klebanov, Nikita Nekrasov; group will write up first |
| Process | Three parallel shepherd experiments | — | Feedback invited; OpenAI notes selection of first shepherds creates information asymmetry |
Spelling note: the conjecture name is Kakeya (OpenAI primary post). Some secondary URLs use other spellings.
Background
In 1991 Sogge proposed local smoothing for waves; Guth–Wang–Zhang proved it in two spatial dimensions (2019), and Wang–Zahl proved Kakeya in three dimensions (2025). OpenAI says its models produced an all-dimension local smoothing proof via a novel auxiliary PDE, obtained October 18, 2026 and Lean-formalized the next day.
ETH, proposed by Impagliazzo, Paturi, and Zane, underpins many conditional hardness results. OpenAI says its system found a natural-language disproof on October 16, 2026 and a Lean formalization on October 17, and that the result reopens how fast some tasks could become—while turning theory into practical algorithms still needs further work. The shepherd group asked that raw data be public and that OpenAI fund a conference on implications.
On physics, OpenAI situates the claim in Balaban-style constructive QFT: existence for interacting theories with gauge fields and fermions in four dimensions, including QCD and a grand-unified completion of the Standard Model, with pure Yang–Mills existence as a subproblem tied to the Clay Millennium Prize statement. It dates natural-language Yang–Mills existence to September 9, 2026, QCD to September 18, and Lean formalization of Yang–Mills existence to October 16, 2026.
Reactions and Impact
The Verge reported the three claims and the shepherd model, quoting OpenAI's GitHub blurbs for local smoothing and the ETH refutation, and noting that physics details remain withheld. The same piece cites an AGMAI panel of math advisers recommending AI labs not adopt practices that designate certain people to receive outputs first (unequal access, borrowed reputation, biased selection), preferring scholarly-repository releases with Lean proofs, problem-selection disclosure, and funding for people who unpack results. OpenAI itself acknowledges tradeoffs: shepherds can organize understanding, but first access is asymmetric.
What We Don't Know Yet
- Whether the Lean formalizations and natural-language proofs will survive independent checking, simplification, and peer review; OpenAI presents them as AI output plus claimed formalization, not as completed community verification.
- Full physics manuscripts and Lean artifacts (deferred to the Gross-led group).
- How credit, early access, and conflict-of-interest norms will settle—questions the ETH shepherd group explicitly opened to the community.
- Practical algorithmic consequences if ETH is false; OpenAI says finding and deploying faster algorithms still requires significant further research.
- Relation to OpenAI's earlier large math dump and its $5 million Simons Foundation contribution for workshops (mentioned as complementary support, not a substitute for review).