Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHacker News AI→ original

OpenAI says its AI disproved the ETH conjecture; the proof was formalized in Lean in one day

OpenAI says its model has disproved the Exponential Time Hypothesis (ETH), the belief that problems like SAT, which asks whether a set of constraints can be satisfied, cannot be solved substantially faster than brute-force search. The refutation was found on October 16, and on October 17 it was written up in Lean, where the compiler checks every step of the proof. This is still the company's claim, and the ETH authors are the ones reviewing it. Until they do, the "cannot be faster" statements in hundreds of papers should be read as "in question."

AI-processed from Hacker News AI; edited by Hamidun News
OpenAI says its AI disproved the ETH conjecture; the proof was formalized in Lean in one day
Source: Hacker News AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHacker News AI ↗
Editor-in-Chief
Same releasea proof of the Kakeya conjecture in all dimensions
Key metric #2The ETH is being reviewed by Impagliazzo, Paturi (UC San Diego) and Williams (MIT)
Key metric #3The ETH proof text and Lean code are open on GitHub

OpenAI has released something unusual: not a new model, but three scientific results that, the company says, its internal systems found. They concern mathematics, computational complexity theory and physics. For developers, the most concrete is the claimed refutation of the ETH. Consider the problem of arranging hundreds of dependencies so that none conflict, or building a schedule that satisfies every constraint. Such tasks belong to the SAT class, and for decades it was assumed that as the number of variables grows, solving time inevitably rises exponentially. That is exactly what the ETH asserts. OpenAI says its system showed the hypothesis is false.

What the ETH is and why refuting it matters

The Exponential Time Hypothesis is not a theorem but a widely held belief among researchers: there is no fundamentally fast algorithm for Boolean formula satisfiability. Its authors are Russell Impagliazzo, Ramamohan Paturi and Francis Zane.

Hundreds of papers are built on this belief. The logic runs like this: "If we could solve our graph, database or bioinformatics problem faster than a certain threshold, the ETH would be false. Since people believe the ETH, it can't be done faster." This is how entire catalogs of conditional lower bounds came to exist in graph theory, computational biology, databases and geometry.

If the refutation holds, those bounds stop being reliable. OpenAI itself cautions that this does not produce ready-made fast algorithms. The refutation only reopens the question of how much faster these problems can be solved. Turning theory into practice, the company says, will require substantial additional work.

The other two results

In mathematics, the company claims a proof of the local smoothing conjecture, proposed by Sogge in 1991. Informally, it says a wave can focus very sharply for a moment, but such peaks are short-lived and on average much weaker than the worst case. Kakeya-type conjectures and Stein's bounds follow from it. Local smoothing had previously been proved in two dimensions, and Kakeya was closed for three in 2025. OpenAI's models, it says, proved the conjecture in every dimension, so both consequences are now settled everywhere. The proof rests on a new auxiliary differential equation.

In physics, the company claims a rigorous proof of the existence of quantum chromodynamics (the theory of quarks and gluons) and of a unified extension of the Standard Model in four dimensions. Part of the task was proving the existence of pure Yang–Mills theory, one of the Clay Institute's Millennium Prize problems.

How this is being checked and why it matters to engineers

The most instructive part of the story is the timeline. The model found the ETH refutation on October 16, 2026, and the next day it was formalized in Lean. The mathematical proof was obtained on October 18 and formalized on the 19th. Physics took longer: the models proved Yang–Mills on September 9 and QCD on September 18, and the Yang–Mills formalization was completed on October 16. The source does not say whether the QCD result was formalized in Lean.

Lean turns a proof into code that a compiler verifies, rather than something that depends on human trust. This addresses the main fear about AI-written mathematics: smooth-sounding reasoning that is wrong. A different question remains: whether the Lean statement exactly matches the original conjecture. External groups are working on that.

Each result has its own review scheme. The mathematical proof was released immediately so that researchers could organize their own scrutiny. For the ETH, the review group is led by Impagliazzo, Paturi and Ryan Williams of MIT. The proof text and Lean code are already on GitHub, and OpenAI has committed to funding a conference on the implications. The physics group is led by David Gross, a 2004 Nobel laureate; it includes Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. They are not yet publishing the AI's material and are writing their own exposition.

The terms of collaboration are as follows: experts sign no non-disclosure agreements, take part voluntarily and unpaid, and speak only for themselves. OpenAI does not claim academic credit for the models' results. The company also acknowledges a drawback: choosing who reviews the work determines who gets access first and creates an information imbalance.

What practitioners should do. First, do not expect algorithms to speed up immediately: claims that "it cannot be faster if the ETH holds" should now be read as conditional. Second, adopt the underlying pattern: a model result that can be checked automatically (tests, formal specifications, Lean) is worth more than a polished text. Third, wait for independent reviews. Until the Impagliazzo and Gross groups have weighed in, these are the developer's claims, not an accepted fact. OpenAI has not disclosed how much computation and how many runs were required.

Story

Key Questions Answered

What the ETH is and why refuting it matters?

The Exponential Time Hypothesis is not a theorem but a widely held belief among researchers: there is no fundamentally fast algorithm for Boolean formula satisfiability. Its authors are Russell Impagliazzo, Ramamohan Paturi and Francis Zane. Hundreds of papers are built on this belief. The logic runs like this: "If we could solve our graph, database or bioinformatics problem faster than a certain threshold, the ETH would be false.

What are the key details on The other two results?

In mathematics, the company claims a proof of the local smoothing conjecture, proposed by Sogge in 1991. Informally, it says a wave can focus very sharply for a moment, but such peaks are short-lived and on average much weaker than the worst case. Kakeya-type conjectures and Stein's bounds follow from it. Local smoothing had previously been proved in two dimensions, and Kakeya was closed for three in 2025.

How this is being checked and why it matters to engineers?

The most instructive part of the story is the timeline. The model found the ETH refutation on October 16, 2026, and the next day it was formalized in Lean. The mathematical proof was obtained on October 18 and formalized on the 19th. Physics took longer: the models proved Yang–Mills on September 9 and QCD on September 18, and the Yang–Mills formalization was completed on October 16. The source does not say whether the QCD result was formalized in Lean.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

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.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…