What OpenAI’s AI Proved About The Exponential Time Hypothesis—and What It Didn’t

🔍 Read the full analysis: What OpenAI’s AI Proved About The Exponential Time Hypothesis—and What It Didn’t on ThorstenMeyerAI.com

OpenAI says an unreleased internal AI model produced a proof of a subexponential-time algorithm for exactly counting solutions to fixed-width k-SAT, a claim that would refute the Exponential Time Hypothesis (ETH) for 3-SAT. The company released a preprint and Lean formalization, but the formal check covers a slightly weaker runtime bound than the paper’s sharpest claim, and independent expert review is still pending.

In its October 26 announcement, OpenAI said its system found the result on October 16 and that a Lean formalization followed on October 17. The preprint, dated October 17, was revised on October 25. OpenAI describes the manuscripts it is releasing as “a form of raw research data,” rather than as finished, independently validated findings.

The computer science preprint claims a deterministic algorithm that exactly counts satisfying assignments for any fixed-width k-SAT formula in subexponential time and polynomial memory. Applied to 3-SAT, that would contradict ETH, which says 3-SAT requires exponential time under a specified formal model. The authors also say the result refutes the stronger Strong Exponential Time Hypothesis and three related hypotheses; those are claims in the preprint, not conclusions from a completed external review.

OpenAI published Lean code alongside the paper. The source account reports 3,977 files and about 710,000 lines, and says Lean checks a weaker bound than the paper’s sharpest runtime result. Formal verification means the encoded proof meets Lean’s rules; it does not, by itself, establish that the paper’s unformalized claim follows or that the result has passed expert scrutiny. The source account says the people reviewing the work include Russell Impagliazzo and Ramamohan Paturi, two of the three original authors of ETH, alongside Ryan Williams.

A Challenge to SAT Hardness Assumptions

ETH is used in theoretical computer science to support lower-bound arguments: researchers can show that certain tasks are hard by connecting them to the presumed difficulty of solving 3-SAT. If the new result survives review, some conclusions that rely on ETH would need to be revisited or restated under different assumptions. That is a potential change to a widely used theoretical tool, not evidence that current software or businesses will immediately solve difficult problems faster.

The claimed algorithm remains subexponential, not polynomial. The source account describes its runtime as roughly 2 raised to n divided by the square root of log n, a form that is still far slower than polynomial time. It also reports that OpenAI said practical improvements would require significant further research. No immediate change to commercial computing performance follows from the claim.

The development also draws attention to a verification process in which an AI system generates a mathematical result, a proof assistant checks a formalized version, and researchers assess the argument. That process could help make complex proofs more auditable. Here, however, the gap between the paper’s sharpest bound and the formalized bound, together with the pending review, marks a limit on what can yet be treated as settled.

Popular right now

Roborock RockNeo Q110H Wire-Free Robot Lawn Mower

AmazonView Latest Price

Breville Barista Pro Espresso Machine with Grinder, Brushed Stainless Steel

AmazonView Latest Price

Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

What ETH Says About 3-SAT

In k-SAT, each rule is a logical clause containing at most k variables or their negations. The question is whether assignments of true or false to the variables can satisfy all the clauses. With n variables, trying every assignment takes 2^n possibilities. Researchers have developed faster-than-brute-force algorithms, but ETH says that for 3-SAT the running time cannot be reduced to subexponential time.

ETH was introduced by Russell Impagliazzo, Ramamohan Paturi and Francis Zane in 2001. It is stronger than the claim P ≠ NP: refuting ETH would not show that every problem whose answers are easy to check is also easy to solve. The Clay Mathematics Institute still lists P vs. NP as unsolved, according to the source material. The source also notes that views on ETH have not been unanimous; it cites Scott Aaronson’s 2017 description of the lack of a consensus that ETH is true.

The ETH announcement was one of three research claims OpenAI described on October 26. The company also announced work in harmonic analysis and physics, with different outside groups involved in reviewing or stewarding those materials. The ETH preprint and code were made public at the request of its external group, according to OpenAI.

The Formal Proof’s Remaining Gaps

The central open question is whether independent researchers agree that the paper’s argument proves the claimed result. The source account says a group led by Impagliazzo, Paturi and Williams will examine it, but reports no completed review or final endorsement. Until that work is done, the announcement remains a claim supported by a public preprint and a formalization, not an independently confirmed refutation of ETH.

There is also a specific difference in scope: the Lean formalization checks a slightly weaker runtime bound than the sharpest one stated in the paper. The available information does not explain whether the stronger bound can be formalized, whether reviewers consider the gap consequential, or whether revisions will close it. OpenAI’s account of the discovery and release process is the source for the dates and collaboration terms; those details are not independent confirmation of the mathematical claim.

Finally, even if the proof is accepted, the information provided does not establish practical performance on real-world SAT instances, an implementation, or a timeline for applications. The result should not be described as solving P vs. NP.

Expert Review and Formal Checks

The next milestone is the external group’s examination of the preprint and Lean code. That review will need to address both the mathematical argument and the difference between the paper’s strongest runtime bound and the weaker bound covered by the formalization. OpenAI says it has agreed to fund a conference on the result, which could provide a venue for further discussion; the source material gives no date for that meeting.

Researchers may then clarify whether the paper’s claims need revision, whether additional formal verification is possible, and what consequences follow for results that rely on ETH. Until those steps are reported, the most accurate description is that OpenAI has announced a potentially significant AI-generated proof, with part of its claim machine-checked and independent validation still to come.

Key Questions

What does OpenAI say its AI proved?

OpenAI says an internal model produced a proof of a deterministic, subexponential-time algorithm for exactly counting solutions to fixed-width k-SAT. For 3-SAT, that would refute the Exponential Time Hypothesis.

Does this settle P vs. NP?

No. The claimed algorithm is still subexponential rather than polynomial, and the result does not resolve whether every problem whose solutions are easy to check is also easy to solve.

Has Lean verified the entire claim?

Not according to the source account. Lean checked a formalized result with a slightly weaker runtime bound than the sharpest bound in the preprint. Independent review is also pending.

Will the result make computers solve SAT problems faster now?

No immediate practical improvement has been established. OpenAI said further research would be needed, and the source material does not report a practical implementation or tested performance gains.

Who is reviewing the result?

The source account says an external group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will examine the work. It reports that no final expert sign-off has yet been issued.

Source: ThorstenMeyerAI.com