# OpenAI uses its AI to bring down ETH and prepares its review

> OpenAI says its models refuted ETH, solved Kakeya and proved Yang-Mills, with outside groups reviewing the result without pay.
> By José Maldonado · Artificial Intelligence · 2026-10-27

Theoretical computer science rests on a conjecture that almost no one questions out loud. This Monday, OpenAI said its models have refuted it, that the proof is written in Lean and that anyone can open the repository and look for the seams. The conjecture is the exponential time hypothesis, and hundreds of papers hang on it, taking for granted where the line between the tractable and the intractable lies.

Its fall would turn the whole world of computer science upside down, which is why everyone is holding their breath. In fact, the same note publishes a result in harmonic analysis that drags the Kakeya conjecture down with it, and another in quantum field theory that grazes a Millennium Problem. All three came out of the model between October 16 and 18, and all three arrived with a Lean formalization and an outside group assigned to verify them. And this time, they are people who are deep in the thick of the field they are seeking to transform.

That is the fundamental change from September, when the same OpenAI announced a solution to Navier-Stokes and the ensuing disorder laid the procedure bare. Now OpenAI's AI arrives with reviewers who are not paid, but above all and more importantly: **with no confidentiality agreements and with the raw data committed**.

## What is ETH and what does it mean for it to fall?

The exponential time hypothesis, or ETH, was born in the late nineties in a paper by Russell Impagliazzo, Ramamohan Paturi and Francis Zane. That text defined the k-SAT problem by its asymptotic cost and conjectured that, for k greater than or equal to three, no subexponential algorithm exists. Put simply: **ETH tells us that in computing, there are no magic solutions that solve problems quickly, other than testing the combinatorics of their possibilities exponentially.** 

A simple way to see this is a combination lock. If the lock has 5 switches, the possible solutions are 2^5 combinations, or in other words, 32 possible combinations. Something a person can crack in a few minutes. But if the lock has 256 (as an AES-256 key might), the total combinations are 1.15792089237^77 combinations, so many that not even in the entire age of the Universe could you solve it.

ETH is widely used in cryptography, to "guarantee" that an algorithm is secure against brute-force attacks, even when it has been weakened. So its breaking will certainly shake foundations that would lead many to question half the world of computer science.

### ETH is much more than security

But ETH also introduced a type of reduction that preserves subexponential complexity and showed that half a dozen classic problems behave the same as satisfiability. From there came a shortcut that allows one to assert that something is intrinsically hard without proving it from scratch.

Half a dozen fields have used that shortcut. Graph theory and computational biology cite it daily, and databases use it to explain why some queries cannot be rushed. All those hardness results rest on the same ground: ETH.

The work OpenAI attributes to its models gives a deterministic algorithm that counts the assignments satisfying fixed-width CNF formulas in subexponential time and polynomial space. The 3-SAT case, the most cited in the family, brings down the conjecture, according to the [company's announcement](https://openai.com/index/building-scientific-understanding-with-academic-communities/). Basically, OpenAI and its AI model have said: ETH is false.

If the result holds, it does not just change a belief: it reopens the question of how much faster those problems can become and where to look for better algorithms. However, we must remember that a theoretical refutation does not deliver a practical algorithm.

What it does is remove a foundation and force a review of which hardness results still stand. No one is going to notice on Monday that their SAT solver runs faster, and that is precisely the misunderstanding worth avoiding.

Result

Model output

Lean formalization

Material released

Refutation of ETH

October 16

October 17

Written and proof in the open

Local smoothing and Kakeya

October 18

October 19

Written and proof in the open

Yang-Mills and QCD in 4D

September 9 and 18

October 16

Retained by the group

## The reviewing group certifies nothing

The interesting part of the announcement is who is going to look at it. The group is led by Russell Impagliazzo and Ramamohan Paturi, of the University of California, San Diego, with Ryan Williams from MIT. The first two proposed the hypothesis alongside Francis Zane almost three decades ago.

**In other words: those reviewing are the authors of the conjecture that is falling.** Williams, for his part, has spent years refuting strong versions of ETH, so he knows the terrain from both sides and knows where the holes usually are.

The conditions of the assignment matter more than the names. No one has asked them to sign a confidentiality agreement, they are not paid and their role is not to endorse the finding. In the [open letter to the field](https://docs.google.com/document/d/1adB0IhJgLzsmbfCgoLrtoJTjnlPf74UnyjMo35P5JHo/edit) the three warn that their involvement certifies nothing.

They also asked for the raw data and funding for a conference where the community can discuss the implications. OpenAI accepted both, and the money has a name and a date: on October 15 it confirmed five million dollars for an initiative of the Simons Foundation.

That fund will allow mathematicians to examine and extend results produced by AI, with access to the models and computing capacity for funded researchers. In the same gesture, the company acknowledged that it had been consulting the advisory group of the Institute for Advanced Study.

One problem remains that the note itself admits without embellishment: choosing the first reviewers conditions who gets in first and creates information asymmetry. It is the part of the process that no press release resolves, because it depends on who is invited.

## Kakeya falls through a side door

The second point under attack is Kakeya, although OpenAI's mathematical result does not attack Kakeya head-on. In 1991, Christopher Sogge proposed that the extreme behavior of a wave is short-lived: the worst that can happen in an instant is far more severe than what is seen when averaging.

From that local smoothing conjecture follow the Kakeya conjecture and the Stein restriction conjecture, which had resisted direct attacks for decades. It is the kind of problem where progress is measured in dimensions conquered, one by one.

The track record helps gauge the leap. The known proof covered two spatial dimensions since 2019, and in 2025 Wang and Zahl solved Kakeya in three dimensions. OpenAI's models claim to have proved local smoothing in all dimensions (N-dimensions).

They achieved it with an auxiliary partial differential equation that no one had used for this specific problem, and if the argument is correct, all three conjectures fall at once in any dimension. The bet is large and the margin for error, narrow.

Here the company chose the most expeditious of the three routes: release the model's output and let the community organize itself, with no prior review or human filter before publishing it. We will surely have news of this in the coming days.

## Yang-Mills, the prize and a missing word

Finally, in physics the result is the most ambitious of all and, if true, would completely transform our view of particle physics, the Standard Model and quantum physics. How? A rigorous proof that certain interacting quantum field theories exist in four dimensions. It includes quantum chromodynamics (QCD), which describes the strong force between quarks and gluons.

As a byproduct, pure Yang-Mills theory appears, and there the announcement grazes one of the seven Millennium Problems without claiming it. The nuance lies in the original statement, which asks for two things and not one.

The [Clay problem](https://www.claymath.org/millennium/yang-mills-the-maths-gap/) requires that a non-trivial quantum Yang-Mills theory exist on four-dimensional space and that it have a positive mass gap. Put briefly: the lightest particle of the field cannot be arbitrarily light.

OpenAI announces that its model has managed to solve this problem. The gap is not mentioned, and that absence is the difference between an enormous step and the million dollars. It is the detail that separates a result for the ages from an optimistic headline.

The reviewing group is headed by David Gross, Nobel laureate in Physics in 2004 for asymptotic freedom. He is joined by Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, names that have been surveying that terrain for two decades.

Jaffe formulated the problem with Witten and co-founded Clay. Jaffe is a towering figure in the world of physics, even more than Stephen Hawking (theoretical physicist); after Jaffe it gets more practical and with more than 200 papers, he makes his role in the community clear. For his part, Douglas has published two status reports on the same subject (Yang-Mills), one in 2004 and another at the beginning of this year. The choice of both is clear, they are the people best prepared to undertake the review of that result.

However, OpenAI has made a decision that contrasts with the other two problems: it has not published the material generated by the AI. . The company also does not claim the prize, just as it said in September it would not for Navier-Stokes.

It is the only one of the three routes where the model's artifact is not on view, and for that very reason the hardest to audit from outside. Anyone who wants to check the result will have to wait for the group to publish its version.

## What broke in September

Here, it is worth recalling why this protocol exists. On September 8, OpenAI announced a solution to the Navier-Stokes problem with a 166-page document produced by an internal model, and the work behind it was not a chat.

A swarm of some 10,000 agents spent 88 hours of compute and 130 billion tokens building the counterexample that makes the equations blow up. It was presented as a continuation of the program of two Spanish mathematicians, Diego Córdoba and Luis Martínez-Zoroa.

That unleashed a public conflict with Tristan Buckmaster and Levent Alpöge, who had spent a year pursuing that route with competitors' models. The account of the cross-accusations, including pressure on one of them not to speak, is in this [review of the authorship scandal](https://observatorioblockchain.com/en/ai-agents/ia-roba-ideas-espanoles-openai/).

Then came the technical problem, which is the one that really matters for what is happening today. Three mathematicians from Cambridge and King's College documented two points where the proof written in natural language did not match the Lean version that was supposed to verify it.

### An uncomfortable cause

After all, a system that formalizes has to produce code that compiles, and when it encounters a passage that does not compile it looks for a way out, even if that means drifting from the text it claims to be translating.

The result can be an impeccable formal proof of a statement weaker than the one in the paper. That is the gap that human review exists to detect and that no automatic verifier flags, as the team detailed in its [analysis of the discrepancies](https://www.newscientist.com/article/2592824-openai-mistranslated-mathematics-into-code-for-its-navier-stokes-proof/).

On October 6 came the next chapter, which serves as a thermometer. OpenAI published 722 preprints with 372 distinct results, the fruit of having tried its luck with some 4,000 problems and an average cost of three hours of reasoning per result.

Weeks later it had withdrawn three manuscripts and revised fourteen, and of the 719 that remained standing only about 300 were formalized in Lean, according to the tally published by [the scientific coverage of the episode](https://www.science.org/content/article/openai-just-changed-math-forever-exactly-how-no-one-yet-knows).

Terence Tao's complaint, gathered in this [assessment of the field's demands](https://techcrunch.com/2026/10/08/openais-math-solutions-arent-meeting-the-fields-standards-yet/), pointed to the heart of the matter. Problems solved by someone who cannot give a talk on the result or answer questions about it.

## What is missing to call it a theorem

In any case, this work is not definitive. What is always missing is what takes longest: people with judgment reading, wrestling with the details and publishing where it breaks. The schedule accepted by OpenAI, with reviewers under no contract, open data and a paid conference, is real progress compared with the September fiasco.

But it does not turn a model output into settled knowledge. Between one thing and the other there are months of human work and some correction that no one can anticipate now, and that holds equally for all three results.

What has changed this week is the bar. A lab that announces three open problems in a single day can no longer ask for borrowed trust: it has to deliver the scaffolding for others to earn it or destroy it.

If Impagliazzo's group finds a hole in the refutation of ETH, we will know exactly where to look, which is more than we had in September. And if it does not, the word theorem will keep waiting its turn.
