OpenAI says AI disproved 1999 ETH hypothesis and hands verification to experts
OpenAI says its model disproved the Exponential Time Hypothesis (ETH), a long-standing assumption that problems like 3-SAT cannot be solved much faster than near-complete brute-force search. It's like a lock with a thousand combinations: the consensus was that there was no way around brute force, yet the AI reportedly found a shortcut. A group of Impagliazzo, Paturi and Williams will carry out the verification, and the Lean proof materials are on GitHub. For now this is the company's claim, and it should be trusted only after external review.
AI-processed from The Verge; edited by Hamidun News
Less than a month after the mass publication of hundreds of manuscripts claiming "solutions to long-standing problems," OpenAI announced three new "major" discoveries: in mathematics, computer science and physics. The one most understandable to an engineer is the claim that the model disproved the Exponential Time Hypothesis (ETH). The hypothesis was proposed in 1999, and many considered it correct. The release is also notable for the company bringing in outside experts, whom it calls "shepherds": they guide the verification. The problem this is trying to address is familiar to anyone who has worked with LLMs: a model produces a convincing long result faster than people can understand whether it is correct.
What ETH is and what OpenAI claimed
Imagine a package manager that selects compatible versions of hundreds of libraries, or a scheduler that arranges tasks under conditions like "if A, then not B." Such puzzles reduce to the classic 3-SAT problem: you must assign true/false values so that a set of logical conditions holds. ETH is the assumption that, in the worst case, there is no way around brute-force search, whose running time grows exponentially with the size of the problem.
According to OpenAI, the model produced a deterministic algorithm that exactly counts the number of satisfying assignments for CNF formulas of fixed width. It runs in subexponential time (that is, noticeably faster than exponential growth) and uses polynomial memory. For the 3-SAT case, the company claims, this disproves ETH. If the conclusion holds up, papers built on this assumption will also need to be revisited.
The two other results look less verifiable so far. In mathematics, OpenAI says the model proved the local smoothing conjecture for waves in every dimension two and higher. It implies the Kakeya conjecture and the Stein restriction conjecture. For context, Hong Wang of New York University received the Fields Medal a few months ago for solving the Kakeya conjecture in three dimensions. Here the model's output was released without edits, as in the October 6 release, and researchers are invited to work through the explanations themselves.
In physics, the claim concerns a "mathematically rigorous proof of existence" for certain quantum field theories with gauge fields and fermions in four dimensions, including quantum chromodynamics (QCD) and the "unified completion" of the Standard Model. The full text has not been published. The work will be accompanied by a group led by David Gross, including Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov.
How the verification works
For computer science, OpenAI named an independent working group: Russell Impagliazzo and Ramamohan Paturi of the University of California, San Diego (who were part of the group that proposed ETH) and MIT professor Ryan Williams. At the group's request, the AI-written text and the Lean proof were posted to GitHub immediately. Lean is a system in which a proof is written so that its steps can be checked by a program. For developers this logic is familiar: a model's result should arrive together with a verifiable artifact, like code with tests, not just a persuasive text.
Alongside the release, the company published an open letter to the community with three questions:
- What rules should govern early access to AI results, disclosure of conflicts of interest, and the use of unpublished materials?
- How should contributions be credited: both for the discovery itself and for verification, correction, simplification and explanation?
- How can students and researchers be supported so that they keep the ability to doubt results and come up with new ones?
Where the approach is disputed
The company itself acknowledges that choosing "shepherds" may cause problems. The AGMAI council of mathematical advisers, which OpenAI assembled itself, wrote the day before: "we recommend that AI companies not apply" the practice of giving results to individual people first. The reasons: unequal access, borrowed reputation and biased selection. The council proposed a different order: publish results in scientific repositories with Lean proofs, disclose how the problems were chosen, and fund those who can assess the results.
So OpenAI is using a format its own advisers had just recommended against. This does not mean the results are wrong, but calling them confirmed is premature: there is the company's claim, files on GitHub and a promised review by independent groups. For physics there is not even a full text.
What practitioners should take away
For those building products on AI, the lesson is simple: the bottleneck has shifted from generation to verification. If a model produces results faster than people can check them, verification becomes the main cost. The material contains no data on compute costs, so the economics of these discoveries cannot be assessed. The most sensible thing is to follow the ETH story: the proof is written in Lean, and this is the first serious test of whether a claimed AI result survives external scrutiny.
- Nov 21, 2026Altman Calls Handing Decisions to AI a "Terrible Idea" as OpenAI Pulls 25% of Engineers Onto Security
- Nov 21, 2026Anthropic shares opened at $111.50, valuing the seven co-founders' stakes at $180 billion
- Nov 21, 2026OpenAI Lets Six Apollo Auditors Inside for Four Months
- Nov 21, 2026Microsoft Releases Seven MAI Models: Coding Model Delivers 98% of GPT-6.1 Sol's Quality at 40% of the Price
Key Questions Answered
What ETH is and what OpenAI claimed?
Imagine a package manager that selects compatible versions of hundreds of libraries, or a scheduler that arranges tasks under conditions like "if A, then not B." Such puzzles reduce to the classic 3-SAT problem: you must assign true/false values so that a set of logical conditions holds. ETH is the assumption that, in the worst case, there is no way around brute-force search, whose running time grows exponentially with the size of the problem.
How the verification works?
For computer science, OpenAI named an independent working group: Russell Impagliazzo and Ramamohan Paturi of the University of California, San Diego (who were part of the group that proposed ETH) and MIT professor Ryan Williams. At the group's request, the AI-written text and the Lean proof were posted to GitHub immediately. Lean is a system in which a proof is written so that its steps can be checked by a program.
Where the approach is disputed?
The company itself acknowledges that choosing "shepherds" may cause problems. The AGMAI council of mathematical advisers, which OpenAI assembled itself, wrote the day before: "we recommend that AI companies not apply" the practice of giving results to individual people first. The reasons: unequal access, borrowed reputation and biased selection.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.