AI Disproves Exponential Time Hypothesis — If Validation Holds

Key Takeaways

• OpenAI announced a proof on 26 October 2026, and Anthropic announced a proof on 27 October 2026, each said to disprove the Exponential Time Hypothesis, according to the briefing source for this story.
• Both results were formalised in Lean, a proof verifier that checks a proof step by step, rather than relying only on human reading.
• The Exponential Time Hypothesis, in plain terms, says that the problem known as 3-SAT cannot be solved much faster than brute force.
• Russell Impagliazzo and Ramamohan Paturi, who formulated the hypothesis in the late 1990s, together with MIT’s Ryan Williams, are leading external validation.
• Review is still ongoing, so this story is best described as announced and under validation, not as settled textbook fact. If confirmed, it would affect the premises of many studies that relied on the hypothesis.

A result announced, not yet a settled fact

The most important sentence in this story is also the most easily lost: review is still ongoing. OpenAI and Anthropic have announced proofs that, if correct, disprove a longstanding premise in computational theory. External experts are checking them. Until that checking is complete, the correct description is that the Exponential Time Hypothesis has been announced as disproved, not that it has been conclusively disproved for all time. That caution is not a lack of enthusiasm. It is how mathematics works, and it is especially important when the claim comes from AI systems, because the stakes are high and the history of claimed breakthroughs includes both genuine revolutions and errors found on closer inspection.

The briefing that summarises the announcements, published by Kim Kyung-jin, puts the same point plainly: the fact that two companies announced the same finding demands rigorous verification rather than boasting, and until external validation concludes, nothing is definitive. That is the frame this article keeps throughout. We explain what was announced, what it would mean if it holds, and what could still go wrong in validation, without treating any of those stages as already finished.

It is also worth separating this story from a different, unrelated project that readers may have seen mentioned elsewhere. This is not about the Axiom project or the Classification of Finite Simple Groups. That is a different theorem, a different project, and a different draft, and it is not linked or discussed here. Confusing the two would be a serious error, because they involve different mathematics, different claims, and different evidence. This article is only about the Exponential Time Hypothesis and the problem called 3-SAT.

What the Exponential Time Hypothesis actually says

Computational theory studies not just whether a problem can be solved, but how quickly. Some problems have fast solutions. Some have solutions that are known, but only if you are willing to wait an impractical amount of time. And some problems sit in a grey area where nobody knows whether a fast solution exists at all. The Exponential Time Hypothesis, usually shortened to ETH, is a precise claim about one of those grey areas.

Start with 3-SAT, because ETH is a claim about 3-SAT. Imagine a puzzle made of many small either-or conditions. Each condition involves three variables, each of which can be true or false, and the puzzle asks whether there is any assignment of true and false to all the variables that satisfies every condition at once. That is 3-SAT in its simplest form: a logic puzzle with three literals per clause, and a yes-or-no question about whether a satisfying assignment exists.

The brute-force way to solve such a puzzle is to try assignments one by one until you find one that works, or until you have tried them all and can say that none works. If there are n variables, there are 2 to the power of n possible assignments. That number grows explosively. Add ten variables, and the number of assignments multiplies by about a thousand. Add twenty, and it multiplies by about a million. For even a modest number of variables, brute force quickly becomes impossible in practice, even for very fast computers.

ETH says, in plain terms, that you cannot do much better than brute force for 3-SAT. More precisely, it says that there is no algorithm that solves 3-SAT in time that is substantially faster than exponential in the number of variables. It does not say that 3-SAT is unsolvable. It says that any solution will, in the worst case, take something close to exponential time. That is a strong claim, because it rules out a whole family of clever shortcuts. If ETH is true, then a large class of problems that can be reduced to 3-SAT inherit the same hardness. If ETH is false, then at least one faster route exists, and the whole structure of assumptions built on ETH has to be re-examined.

That is why the hypothesis matters beyond a single puzzle. Theoretical computer scientists have used ETH as a foundation for hundreds of studies, proving that other problems are hard by showing that a fast solution to them would imply a fast solution to 3-SAT, which ETH says cannot exist. Remove the foundation, and those proofs do not automatically become false, but their logic has to be rebuilt on different ground. Some results may survive with new arguments. Others may turn out to be weaker than previously believed. The briefing source makes exactly this point: if the disproof is confirmed, the premises of hundreds of studies relying on this hypothesis will be upended.

What OpenAI and Anthropic announced

According to the briefing source, OpenAI announced its proof on 26 October 2026, and Anthropic announced its proof a day later, on 27 October 2026. Each company said it had used its own AI model to produce a proof disproving ETH. The briefing also reports that, on 16 October, both models solved the problem with the identical bound expression, written as 2 raised to the power of n divided by the square root of log n, and that the results were formalised using Lean.

That bound expression deserves a plain explanation, because it is the heart of the claim. The expression 2^(n/√log n) grows more slowly than 2^n, the brute-force bound, but it still grows faster than any polynomial in n. In other words, the announced algorithm is not claiming to solve 3-SAT quickly in the everyday sense. It is claiming to solve it noticeably faster than brute force, and fast enough to break the specific barrier that ETH sets. In complexity theory, that distinction is everything. ETH does not forbid every improvement over brute force. It forbids improvements beyond a certain threshold. Crossing that threshold, even by the amount described here, is enough to disprove the hypothesis as stated.

Anthropic’s paper is titled “A 2^o(n) Time Algorithm for 3-SAT”, according to the briefing source. The notation 2^o(n) is a compact way of saying that the running time grows as 2 to a power that is much smaller than n itself, which is the formal way of stating that the algorithm beats the exponential barrier ETH imposes. Readers do not need to follow the notation in detail to grasp the point: the claim is not that 3-SAT is now easy, but that it is not as hard as ETH said it was.

The fact that two independent teams, using different models, arrived at the same bound is striking, and it is one reason the announcements attracted so much attention. It is also a reason for caution, not for certainty. Two teams arriving at the same answer could mean that the answer is correct and robust. It could also mean that both teams were working from shared ideas, shared literature, or similar model training, and that a common error or a common hidden assumption is present in both. Independent validation by people who did not produce the proofs is the only way to distinguish those possibilities, which is why the role of the external validators matters so much.

Why Lean matters, and what it does not settle

Both proofs were formalised in Lean, a proof verifier. That detail is important, and it is often misunderstood. Lean is not an AI model that says “this looks right”. It is a formal system that checks a proof mechanically, step by step, against a set of logical rules. If a proof is written in Lean’s language and Lean accepts it, then every step follows from the previous steps according to those rules, with no gaps that a human reader might overlook. That is a much stronger check than a human skimming a long argument and finding it plausible.

At the same time, Lean does not settle everything. A Lean proof is only as good as the statement it proves. If the formal statement in Lean does not exactly match the mathematical claim being announced, then Lean can accept a correct proof of a slightly different statement while the original claim remains unproven. Similarly, Lean checks logical steps, but it does not check whether the definitions used in the formalisation capture the intended mathematical objects. Those are exactly the kinds of issues that external validation is designed to catch: reading the formalisation, checking that the statement matches ETH as formulated, and confirming that no subtle mismatch has crept in.

There is also a human dimension. A very long, machine-checked proof can be correct and still be difficult for humans to understand. That does not make it wrong, but it does mean that the mathematical community will want time to digest it, to find a more human-readable explanation, and to check whether the ideas in the proof can be used elsewhere. Validation is therefore not a single moment when someone says “yes”. It is a process, and the briefing source is clear that the process is still ongoing.

Who is checking, and why their names matter

The external validation is being led by Russell Impagliazzo and Ramamohan Paturi, who originally formulated the hypothesis in the late 1990s, together with MIT’s Ryan Williams, according to the briefing source. That is a significant detail, because the people best placed to judge whether ETH has been disproved are the people who stated it most precisely in the first place. They know what the hypothesis was intended to claim, what its edge cases are, and what would count as a genuine counterexample rather than a proof of a weaker or different statement.

Their involvement also cuts against the simplest sceptical reading, which is that the announcements are mere publicity. If the formulators of the hypothesis are willing to lead the checking, then the claim is being taken seriously by the people with the most to lose if it is wrong and the most authority to say whether it is right. At the same time, their involvement does not pre-judge the outcome. Leading validation means checking carefully, and careful checking can end in either confirmation or rejection. Readers should resist the temptation to treat the validators’ participation as a verdict before they have delivered one.

Why a disproof would matter beyond theory

If the disproof is confirmed, the first effect will be inside theoretical computer science itself. Researchers will need to revisit results that assumed ETH, to see which ones still hold, which ones need new proofs, and which ones were actually stronger or weaker than believed. That is normal scientific work, and it is not a sign of failure. Hypotheses are meant to be tested, and a field that can absorb the loss of a major hypothesis and rebuild on firmer ground is a healthy field.

The second effect is about AI and mathematics. These proofs were produced with AI models, and formalised in a machine-checkable system. That combination, AI generating a candidate proof and a verifier checking it, is a powerful workflow, and it is already changing how mathematical research is done. A confirmed disproof of ETH by this route would be a landmark example, not because AI replaced mathematicians, but because it extended what mathematicians could attempt. Human experts still framed the problem, checked the formalisation, and will judge the result. The AI did not work alone, and the story should not be told as if it did.

The third effect is more practical, and it should be stated carefully. 3-SAT is related to a wide family of problems in scheduling, verification, planning, and optimisation, where the task is to find an assignment that satisfies many constraints at once. A faster algorithm for 3-SAT does not automatically make those practical problems easy, because the announced bound is still exponential, just less steeply exponential than brute force. For large real-world instances, the difference between 2^n and 2^(n/√log n) may still leave the problem out of reach. What changes is the theoretical landscape, and, over time, the ideas in the proof may lead to better practical algorithms for specific kinds of instances. That is a possibility, not a promise, and it should not be oversold.

For readers who want a gentler introduction to how AI systems handle structured knowledge and reasoning, our explainer What Is RAG? How AI Answers From Your Own Documents covers a different but related idea: how models combine retrieved information with generation. And for a broader view of where frontier AI may be heading, see AI to SI: From Today’s AI to Tomorrow’s Superintelligence. Neither is a substitute for the technical proof, but both help place this announcement in the wider story of AI capability. A useful third perspective is Andrej Karpathy on the AI Adoption Gap: Why Frontier AI Feels Miles Ahead of Everyday Use, which reminds us that a frontier breakthrough and everyday usefulness are not the same thing.

What could still go wrong

Validation can fail in several ways, and naming them helps readers understand why patience is warranted. The formal statement could turn out to differ from ETH as originally formulated. A definition could be subtly off, so that the proof covers a variant of 3-SAT rather than 3-SAT itself. An assumption buried in a long proof could turn out to be unjustified, even if Lean accepted the steps that follow from it, because Lean checks derivations, not the truth of every starting assumption in the way a human might intend. Or the proof could be correct but rest on a reading of ETH that other theorists dispute, leading to a debate about what was actually disproved.

None of these possibilities is a prediction that the proof is wrong. They are the standard reasons why mathematics insists on independent checking, especially for a result of this size. The history of mathematics contains famous claimed proofs that took years to verify, and others that were found to contain gaps. The system worked in both cases because it did not rush to declare victory. The same discipline applies here, and it applies with extra force because AI-generated proofs are new enough that the community is still learning how best to check them.

What to watch next

The next milestones are not new announcements, but the quiet work of validation. Watch for statements from Impagliazzo, Paturi, and Williams, and for any published notes that explain the proof in human-readable form. Watch for whether other research groups independently reproduce the formalisation, and whether the bound expression holds up under scrutiny. Watch, too, for the first papers that build on the result, assuming it holds, and for the first papers that identify which ETH-based results survive and which do not. Those follow-up papers will be the real measure of the disproof’s impact, more than the initial headlines.

For non-specialists, the takeaway is simpler. A major hypothesis in computer science may have fallen, with the help of AI, and the people who created the hypothesis are checking the fall. That is a remarkable story even before it is confirmed, because it shows how far AI-assisted mathematics has come. But the story is not finished, and the responsible way to tell it, at the start of November 2026, is to say exactly that: announced, formalised in Lean, under validation by its own formulators, and awaiting a verdict that only careful checking can give.

Sources: Kim Kyung-jin, AI Daily Briefing 2026-11-02; OpenAI announces proof that the Exponential Time Hypothesis is false.