.svg)
Image: OpenAI logo (Wikimedia Commons) — illustrative image for OpenAI's 26 October 2026 research announcement.
TL;DR — openai exponential time hypothesis disproof explained: On 26 October 2026, OpenAI announced that its internal AI models produced three claimed proofs in mathematics, computer science and physics — including a disproof of the Exponential Time Hypothesis, a conjecture hundreds of papers rely on. Two results come with computer-checked Lean formalisations; none is peer-reviewed yet, and outside expert groups are now examining them. Treat this as a major claim under verification, not a settled result.
📌 KEY HIGHLIGHTS
- OpenAI announced three AI-produced results on 26 October 2026: local smoothing/Kakeya in every dimension, a disproof of the Exponential Time Hypothesis (ETH), and an existence proof for quantum chromodynamics (QCD) and pure Yang–Mills theory.
- The ETH claim says a deterministic machine can exactly count satisfying assignments of fixed-width Boolean formulas in subexponential time and polynomial space.
- The math and computer science results are published with Lean formal proofs on OpenAI's public GitHub; the physics artifacts are being held by the expert group checking them.
- ETH's own proposers — Russell Impagliazzo and Ramamohan Paturi of UC San Diego — plus MIT's Ryan Williams will lead the independent examination of the ETH claim.
- OpenAI says it seeks no academic credit for the results, and warns that turning theory into practical speedups needs significant further research.
What happened?
OpenAI announced on 26 October 2026 that its internal models — not a named public system — produced three research results it describes as raw research data rather than finished science. The company handed each result to a group of outside experts it calls "shepherds", who decide how the material is released, checked and explained. The shepherds work unpaid, with no non-disclosure agreements, and their conclusions are their own.
The announcement matters because of what is claimed, not just who claimed it. The Exponential Time Hypothesis has been a working assumption of theoretical computer science since Impagliazzo, Paturi and Zane proposed it in 1999. Local smoothing and Kakeya are decades-old problems in mathematical analysis. And Yang–Mills existence is part of a Clay Millennium Prize problem carrying a $1 million award. If all three claims survive scrutiny, they would rank among the most consequential results ever attributed to an AI system — which is exactly why scrutiny, not celebration, is the right first response.
Openai exponential time hypothesis disproof explained: why ETH matters
The Exponential Time Hypothesis says that Boolean satisfiability — deciding whether a logical formula can be made true — needs exponential time, and cannot be solved much faster than trying possibilities. Hundreds of papers in graph theory, computational biology, databases and computational geometry derive "hardness" results from it: if a problem is at least as hard as satisfiability, ETH says it cannot be solved fast either.
OpenAI says its system produced the disproof on 16 October 2026 and that it was formalised in Lean on 17 October. According to the formalisation's documentation, the Lean proof establishes that for fixed k of at least 2, a deterministic machine exactly counts the satisfying assignments of a canonically encoded k-CNF formula on n variables in subexponential time and polynomial space. Applied to 3-SAT, the claimed running time is about exp(n/(log n) to the power 1/2+o(1)) — faster than exponential, but still far from fast in any everyday sense.
Two clarifications matter. First, this does not settle P versus NP: a subexponential algorithm is not a polynomial-time algorithm. Second, a Lean formalisation checks that the formal proof steps are logically valid — it does not by itself mean mathematicians accept the result, which is why the independent examination matters. OpenAI itself says practical improvements will require significant further research.
| Result | What OpenAI claims | Status on 27 Oct 2026 | Shepherd group |
|---|---|---|---|
| Local smoothing / Kakeya | Proof in every dimension, resolving Kakeya and Stein's restriction conjectures | AI writeup + Lean proof public on GitHub | Open release; researchers organise their own explanatory work |
| Exponential Time Hypothesis | ETH is false: subexponential counting for fixed-width formulas | AI writeup + Lean proof public on GitHub | Impagliazzo, Paturi (UC San Diego) and Williams (MIT) |
| QCD / Yang–Mills existence | Rigorous existence proof for 4D theories including QCD and pure Yang–Mills | Artifacts held by the physics group; writeup pending | Nobel laureate David Gross, with Douglas, Jaffe, Klebanov and Nekrasov |
Image: Unsplash — illustrative mathematics image (no official product image exists for a proof announcement).
The other two proofs, in plain language
The mathematics result concerns how waves focus. Proposed by Christopher Sogge in 1991, the local smoothing conjecture says the most extreme focusing of waves can only be short-lived. It was proved in two dimensions in 2019, and the three-dimensional Kakeya conjecture was settled in 2025. OpenAI says its models proved local smoothing in every dimension, using what it calls a novel auxiliary partial differential equation — which would also settle the Kakeya and Stein restriction conjectures in all dimensions. The AI output arrived on 18 October and was formalised in Lean the next day.
The physics result is the most carefully caveated. OpenAI says its models produced a mathematically rigorous existence proof for certain interacting quantum field theories in four dimensions, including quantum chromodynamics — the theory of the strong force holding atomic nuclei together — and pure Yang–Mills theory, built on Balaban's constructive field theory approach. But the Clay problem as stated also asks for a mass gap, and OpenAI does not claim that. The Yang–Mills proof arrived on 9 September, the QCD proof on 18 September, and the Lean formalisation of Yang–Mills finished on 16 October. The physics group chose to hold the AI artifacts back and write the result up itself — so unlike the other two, it cannot yet be publicly inspected.
Update 31 October: Anthropic's Claude claims the same disproof — and the hypothesis's own authors respond
Hours after OpenAI's announcement on 26 October 2026, Anthropic posted its own paper, titled "A 2o(n) Time Algorithm for 3-SAT", credited simply to "Claude". Anthropic researcher Levent Alpoge said an internal Claude research model found the core method on 16 October 2026 — the same calendar date OpenAI's model found its disproof — using 14 million output tokens, with no human input in the discovery itself. Both labs say their models independently produced deterministic subexponential algorithms for counting the satisfying assignments of fixed-width Boolean formulas, and both published machine-checked Lean formalisations. Neither result has been peer-reviewed.
| Claim | OpenAI (26 Oct 2026) | Anthropic / Claude (26 Oct 2026) |
|---|---|---|
| Paper | "The Exponential Time Hypothesis Is False" (67 pages, dated 17 Oct, revised 25 Oct) | "A 2o(n) Time Algorithm for 3-SAT" (49 pages, dated 26 Oct) |
| Core discovery | Internal model, 16 Oct; Lean formalisation 17 Oct | Internal Claude research model, 16 Oct; 14 million output tokens, no human input in discovery |
| Claimed runtime | About 2n/(log n)1/2+o(1), polynomial space | 2O(n log log n / √log n) for #3-SAT |
| Formal check | Lean proof published on GitHub | Lean 4 formalisation on GitHub, checked by two separate checkers |
| Review status | Independent examination led by Impagliazzo, Paturi and Williams | No outside review group announced yet; not peer-reviewed |
Two cautions keep the headlines honest. First, a Lean formalisation checks that the formal proof steps follow logically from the definitions — it does not by itself certify that those definitions match the informal problem, and no human had audited Anthropic's 1,050 Lean modules at the time of release. Second, the practical impact is close to zero: Anthropic's paper notes its bound beats brute force only at enormous sizes — for formulas with five clauses per variable, only above roughly 1010,000 variables. It does not break encryption, does not speed up today's solvers, and does not settle P versus NP.
On 31 October 2026, AFP published reactions from the three researchers now organising the independent review — and they are strikingly different. Russell Impagliazzo, who proposed ETH with Ramamohan Paturi in 1999, said: "I would have much preferred it to be discovered by a human who could explain their thought process. And I'd like to have shaken the (actual) hand of the person who disproved ETH." Paturi said his first reaction was surprise, but "the research built around it would not simply disappear", while warning that concentrated access to powerful research tools "raises serious questions about who can participate". Ryan Williams of MIT, a long-time public skeptic of ETH, said he felt vindicated and called it "a major revolution in the field" — while agreeing the result is still theoretical, with "no immediate consequences for almost all people", though future extensions "might require stronger encryption".
What this update changes: the story is no longer one lab's claim. Two competing labs now claim independent AI disproofs of the same 27-year-old hypothesis, found on the same day — and the hypothesis's own authors are leading the check. Until that examination reports, the correct status remains "claimed, Lean-checked, not yet accepted".
Sources for this update: AFP via Malay Mail — scientists' reactions, 31 Oct 2026; TechTimes — two AI systems, same day, 27 Oct 2026; Let's Data Science — OpenAI and Anthropic ETH claims compared.
Why it matters for India readers
For India's large computer science and mathematics community — students preparing for GATE, researchers at IITs and IISc, and engineers working on optimisation, databases and chip design — ETH underpins how hardness is taught and argued. If the disproof holds, textbooks and hundreds of conditional results need re-examination, and some problems long assumed intractable may admit faster algorithms than anyone expected. But the honest position today is "claim under examination": the very people who proposed ETH are leading the check, and they have explicitly said their involvement is not certification of the claim.
There is also a cautionary thread worth noting. Observers have pointed out that the physics paper has not been released, the mass gap is not claimed, and at least one mathematician has noted related human work on 4D Yang–Mills existence via the same Balaban approach. None of that refutes OpenAI's claims — but it is why this post says "claims" throughout, and why the verification process deserves as much attention as the announcement.
Should you care? ✅ Worth following closely, not celebrating yet
Verdict: ✅ Follow this story: if verified, the ETH disproof rewrites a chapter of theoretical computer science, and the Kakeya and Yang–Mills claims are historic on their own. ⏳ But do not treat any of the three as settled — none is peer-reviewed, the physics artifacts are not public, and the examiners themselves stress that verification is under way. Watch the shepherd groups' conclusions, not the headlines. Related reading: OpenAI's GPT-6.1 Terra release, OpenAI's frontier model training without internet, and Claude Mythos 5's AI-designed quantum circuits.
FAQs
What is the Exponential Time Hypothesis?
It is a 1999 conjecture that Boolean satisfiability (SAT) cannot be solved in subexponential time. Hundreds of papers assume it to prove that other problems are hard. OpenAI now claims its AI found a subexponential algorithm, which would make the hypothesis false.
Does disproving ETH mean P equals NP?
No. P versus NP asks whether satisfiability has a polynomial-time algorithm. A subexponential algorithm is faster than exponential but still much slower than polynomial, so P versus NP remains open either way.
Are these proofs verified?
Partially. The math and computer science results have Lean formalisations, meaning a proof assistant checked the formal steps, and the writeups are public on GitHub. But none of the three is peer-reviewed, and the physics artifacts are withheld pending the expert group's own writeup.
What happens next?
Independent examination. The ETH group led by Impagliazzo, Paturi and Williams has invited the community to help organise an open examination, and OpenAI has committed to funding a conference on the work's implications. Expect scrutiny first, verdicts later.
Takeaway: OpenAI's 26 October announcement is the boldest AI-science claim yet — three landmark proofs at once. The evidence so far is a public announcement plus two Lean-checked formalisations, and the real verdict belongs to the independent experts now examining them.
Follow World Tech on TechByCountry for tomorrow's update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Sources
Unite.AI — OpenAI Says Its Models Produced Three Proofs in Math, CS, and Physics (26 Oct 2026 announcement coverage)
DataStudios — OpenAI claims AI disproof of the Exponential Time Hypothesis plus Kakeya and Yang–Mills proofs
CellCog — OpenAI Claims Yang-Mills, ETH and Kakeya Proofs
AI Brief — OpenAI Says Its AI Model Refuted the Exponential Time Hypothesis
Comments
Post a Comment