OpenAI’s AI Research: A New Result, Not A Settled Hypothesis?

🔍 Read the full analysis: OpenAI’s AI Research: A New Result, Not A Settled Hypothesis? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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 on October 26 that an unreleased internal AI model had produced a result contradicting the Exponential Time Hypothesis (ETH), a conjecture that certain Boolean satisfiability problems cannot be solved in subexponential time. The company released a preprint and a Lean proof-checking formalization, but the result has not yet been independently reviewed, and the formalization covers a weaker runtime bound than the paper’s sharpest claim.

The preprint’s central claim is an algorithm that exactly counts the solutions to any k-SAT formula, for fixed k, in subexponential time and polynomial memory. For 3-SAT, the authors say this contradicts ETH. The paper also claims consequences for the stronger Strong Exponential Time Hypothesis (SETH) and three related hypotheses. These are claims in a preprint, not conclusions endorsed by an independent review panel.

OpenAI says its system found the disproof on October 16, and that the Lean formalization followed the next day. The preprint is dated October 17 and was revised October 25. The source account reports that the public Lean release contains 3,977 files and about 710,000 lines; that count is its own tally. Lean verifies that a formal proof follows from its encoded definitions and rules. According to the account, however, the formal statement establishes a weaker bound than the paper’s best runtime result, leaving the sharpest claim outside the formalized proof.

OpenAI has asked an outside group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams to examine the computer science result. Impagliazzo and Paturi were among ETH’s original authors. OpenAI describes its released manuscripts as “a form of raw research data” and says the external researchers will explain and verify the work. The preprint and Lean code are public at the group’s request; review has not been reported as complete.

What an ETH Refutation Would Change

ETH is used in theoretical computer science to support claims that many algorithms cannot run substantially faster than exponential-time methods. If the paper’s result survives expert scrutiny, those arguments would need to be reconsidered where they rely on ETH. That would matter to researchers assessing algorithmic limits, though the practical effects are not established by the announcement.

The stated algorithm runs roughly in 2^(n / √log n) time, according to the source account. This is subexponential, but still much slower than polynomial time. OpenAI has said practical improvements “will require significant further research,” so the result does not mean ordinary software or commercial optimization problems will become faster immediately.

The result is also not a resolution of P versus NP. ETH is a stronger hypothesis about the time needed for specific problems; refuting it does not show that every problem whose answer can be checked quickly can also be solved quickly. The Clay Mathematics Institute continues to list P versus NP as unsolved. For readers outside theory, the immediate significance is chiefly methodological: an AI-generated proof is being presented alongside a machine-checkable formalization and a planned human review, while important gaps between the paper and the checked statement remain.

Products Worth Considering

Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor

AmazonView Latest Price

Jackery Solar Generator 1000 v2 with 200W Solar Panel

AmazonView Latest Price

Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

The Hypothesis Behind the Claim

k-SAT asks whether Boolean variables can be assigned true or false so that every rule in a formula is satisfied; each rule contains at most k variables or their negations. A formula with n variables has 2^n possible assignments, though algorithms can do better than checking each one. The question behind ETH is whether 3-SAT can be solved in time that is subexponential in n.

In 2001, Impagliazzo, Paturi and Francis Zane proposed ETH as a way to express a broader expectation: that 3-SAT still requires exponential time, even after improvements over brute force. The supplied account notes that earlier algorithms reduced the exponential base without removing the exponential form. It also cautions against treating ETH as unanimous belief: a survey by Scott Aaronson described it as an ambitious strengthening of P ≠ NP and said there was no consensus in the field that ETH was true.

The announcement is part of OpenAI’s wider effort to release AI-produced research with outside academic involvement. The company announced three results on October 26: the ETH claim, a mathematics manuscript concerning local smoothing and related conjectures, and a physics claim involving quantum chromodynamics and a grand unified completion of the Standard Model. Their release arrangements differ. The source account also says OpenAI published more than 700 AI-generated manuscripts on October 6, then withdrew three and revised 14 on October 7 after a sign error was identified. That record gives added weight to formal checks and independent review, without itself deciding whether this new proof is correct.

“there’s by no means a consensus in the field that ETH is true”

— Scott Aaronson, as quoted in the supplied source account

Proof Scope and Expert Review

The central uncertainty is whether the paper’s argument is correct and whether its implications follow as stated. The outside group’s examination is pending, and no independent conclusion is included in the supplied material. The Lean formalization offers a machine-checked result, but the reported difference between its weaker bound and the paper’s sharpest runtime claim means it does not formally certify every part of the headline result.

It is also not yet clear what changes, if any, will be made after review, how long the examination will take, or whether reviewers will endorse the claimed refutations of SETH and the other hypotheses. OpenAI’s account says the collaborators are unpaid, sign no non-disclosure agreements and speak for themselves; those arrangements do not substitute for a published assessment of the proof.

Finally, the source material does not establish a practical application, a faster solver for real-world SAT instances, or a change to P versus NP. The runtime bound remains superpolynomial, and OpenAI says further work would be needed for practical gains.

Review, Revisions and Follow-Up

The next milestone is the external group’s review of the computer science manuscript and its formalization. The supplied account says OpenAI has agreed to fund a conference on the result, which could provide a venue for researchers to examine the argument and discuss its scope. No date for the review’s conclusion or the conference is provided.

Readers should watch for a public report from the reviewers, revisions to the preprint or Lean files, and clarification of which runtime claims are formally covered. Until those appear, the accurate description is that OpenAI has announced an AI-produced proof that it says refutes ETH, with a machine-checked weaker bound and independent verification still in progress.

Key Questions

What does OpenAI say its AI proved?

OpenAI says an internal model found an algorithm for exactly counting solutions to fixed-k SAT formulas in subexponential time and polynomial memory. For 3-SAT, that would contradict ETH if the claim is correct.

Has the result been independently confirmed?

No independent conclusion is reported in the supplied material. An outside group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams is expected to examine the work.

Does the Lean proof check the paper’s strongest claim?

Not according to the source account. It says Lean checks a weaker runtime bound, while the preprint’s sharpest bound is not covered by the formalization.

Does this solve P versus NP?

No. The claimed result concerns ETH, a hypothesis about the time required for particular satisfiability problems. P versus NP remains open.

Will this make computers faster right away?

There is no evidence of an immediate practical speed-up. The claimed algorithm remains slower than any polynomial-time algorithm, and OpenAI says practical improvements would require further research.

Source: ThorstenMeyerAI.com