OpenAI’s AI Research Raises A Question About The Exponential Time Hypothesis

🔍 Read the full analysis: OpenAI’s AI Research Raises A Question About The Exponential Time Hypothesis on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

OpenAI said October 26 that an unreleased AI model had produced a proof refuting the Exponential Time Hypothesis (ETH), a long-standing conjecture about how quickly certain hard problems can be solved. The company released a preprint and a Lean formalization, but the result has not yet undergone independent expert review, and the formalized result covers a weaker runtime bound than the paper’s sharpest claim.

The preprint claims a deterministic algorithm can exactly count the satisfying assignments of any k-SAT formula, for fixed k, in subexponential time and polynomial memory. For 3-SAT, the authors say this contradicts ETH, which holds that solving 3-SAT requires time that remains exponential in the number of variables. The preprint also claims to refute the stronger Strong Exponential Time Hypothesis and three related hypotheses.

OpenAI says its system found the disproof on October 16, with a Lean formalization produced the following day. The company released the preprint and code after consulting external academic experts, whom it calls “shepherds.” A review group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams is expected to examine the computer-science result. Impagliazzo and Paturi were among ETH’s original authors.

The formalization contains thousands of files and hundreds of thousands of lines of Lean code, according to the source account. Lean can check that a formal proof follows from its stated assumptions, but the formal statement here checks a slightly weaker runtime bound than the strongest bound claimed in the preprint. That distinction, and whether the underlying argument is sound and significant, remain subjects for expert scrutiny.

What an ETH Refutation Would Change

ETH is used to support conditional lower bounds: results showing that particular algorithms are unlikely to run substantially faster, assuming the hypothesis holds. If the preprint’s claim survives review, researchers would need to revisit those conclusions and identify which bounds depend on ETH or related assumptions. The result could alter theoretical expectations across parts of computer science, although it would not by itself make existing software or commercial systems faster.

The development also tests a possible research workflow: an AI system proposes a mathematical argument, a proof assistant checks a formalized version, and outside specialists assess the result. Formal checking is meaningful evidence that the encoded proof meets Lean’s rules; it does not independently establish that the formalization captures every claim in the paper, that the assumptions are appropriate, or that the result has been interpreted correctly. The independent review is therefore central to how much confidence the wider field should place in the claim.

OpenAI’s announcement says practical improvements would require substantial further research. The claimed algorithm remains subexponential rather than polynomial, and the source account says its speed-up grows very slowly with problem size. No immediate practical consequences have been established.

Products Worth Considering

Samsung 32-Inch Odyssey OLED G8 (G80SH) 4K Gaming Monitor

AmazonView Latest Price

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

AmazonView Latest Price

DJI Osmo Mobile 8 3-Axis Smartphone Gimbal Stabilizer

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

The Conjecture Behind the Claim

In k-SAT, each rule is a logical clause containing at most k variables or their negations. The task is to determine whether assignments of true or false values can satisfy all clauses; counting asks how many such assignments work. With n variables, checking all possible assignments takes 2^n possibilities, though algorithms can improve on brute force while still taking exponential time.

ETH, proposed in 2001 by Impagliazzo, Paturi and Francis Zane, conjectures that 3-SAT cannot be solved in subexponential time. It is stronger than the claim P ≠ NP: refuting ETH would not show that every problem whose solutions are easy to check can also be solved efficiently. The P vs. NP question remains open.

OpenAI’s October 26 announcement also described work in mathematics and physics, but those are separate claims with separate review arrangements. The company said external groups would steward each project and that the physics materials would be withheld at the group’s request. The ETH preprint and formalization were made public. This follows an October 6 release of more than 700 AI-generated manuscripts; the source account says some were later withdrawn or revised after errors were found, underscoring why publication and checking do not replace review.

Review and Formalization Gaps

The central open question is whether independent specialists confirm the preprint’s argument. The review group’s examination has been announced, but no completed assessment or endorsement is reported. Until that work is done, the result should be described as OpenAI’s claim, not as a settled change in computer science.

There is also a specific gap between the paper and its machine-checked version: Lean formalizes a weaker runtime bound than the preprint’s sharpest claim. The source account does not establish whether the stronger bound will later be formalized, or how reviewers will assess the relationship between the formal proof and the full paper. The exact practical value of the claimed algorithm is also unknown; OpenAI says further research would be needed.

Expert Review and Formal Checks

The announced next step is the external group’s review, led by Impagliazzo, Paturi and Williams. Its assessment should clarify whether the proof refutes ETH as stated, how the verified Lean result relates to the paper’s stronger runtime claim, and what qualifications are needed. OpenAI has also said it agreed to fund a conference on the result.

Until reviewers report their findings, readers can distinguish three stages: the AI-generated argument and preprint have been released; a Lean proof checks a narrower formal statement; and independent validation remains pending. Any later account of the claim should be updated to reflect the review’s conclusions rather than treating the announcement alone as final verification.

Key Questions

What does OpenAI claim its AI proved?

OpenAI says its model produced a proof of a deterministic subexponential-time algorithm for exactly counting solutions to fixed-k SAT formulas. For 3-SAT, that would contradict ETH, which conjectures that solving the problem requires exponential time.

Has the result been independently verified?

Not according to the source material. OpenAI released a preprint and Lean formalization and named an external group to review the work, but that review has not yet been reported as complete.

What did Lean check?

Lean checked a formalized proof of a slightly weaker runtime bound than the preprint’s sharpest claim. The distinction matters: the machine-checked statement does not, on its own, establish every part of the paper’s strongest result.

Does this resolve P vs. NP?

No. ETH is a stronger hypothesis about the difficulty of 3-SAT. Even if ETH is refuted, that does not prove P = NP; the P vs. NP problem remains open.

Will this make practical software faster now?

No immediate practical improvement has been established. The claimed algorithm is still subexponential, and OpenAI said practical gains would require significant further research.

Source: ThorstenMeyerAI.com