∇ProvenDecisions
Ctrl K

OpenAI's mass release of machine-written mathematics: a quasi-Riemann claim and a disputed ETH disproof

The month in AI and mathematics

October was dominated by one lab's output. OpenAI says an internal model wrote hundreds of research manuscripts. Many come with Lean proofs, and some of the claims are very large. The main question shifted from "can AI prove things?" to "who checks all of this, and how?" Lean verification carries much of the weight, but it only works if the formal statement matches the claim. Terence Tao wrote on exactly this point. Almost all of what follows is a claim that has not been peer reviewed. Some of it has already been corrected.

OpenAI publishes 722 manuscripts, with a quasi-Riemann result

On 6 October OpenAI put manuscripts from an unreleased internal model into a public repository, openai/math, under an Apache-2.0 licence. Coverage describes 722 manuscripts in 372 related result families. Many have Lean formalizations, and ten come with abridged summaries of the model's reasoning. The OpenAI community mirror of the post is a secondary source for the wording. Counts differ between outlets, so treat the exact numbers with care.

The headline item is a "quasi-Riemann hypothesis" result. Quanta Magazine reports that human experts have unpacked this proof, unlike most earlier AI proofs. Stanford's Jared Duker Lichtman called it the most significant AI-for-math achievement so far. As summarized in secondary coverage, the repository describes a 7/8 zero-free half-plane for Dirichlet L-functions, which include the Riemann zeta function. In plain terms, the result puts a boundary on where certain special functions can have zeros. It does not prove the Riemann hypothesis. That would put all non-trivial zeros on the line with real part 1/2.

What this does not show yet: Quanta's coverage says experts largely agree the shorter proof is correct, but it has not been formally verified. Reports also say one manuscript was withdrawn the day after the release, together with two that depended on it, because of a sign error. That is a useful reminder that a large batch will contain mistakes. The company has said the release is the start of human understanding of these results, not the end. Releasing hundreds of proofs also does not show the model is reliable across mathematics.

A claimed disproof of the Exponential Time Hypothesis

In late October OpenAI announced a preprint titled "The Exponential Time Hypothesis is False". Coverage describes it as a deterministic algorithm that counts satisfying assignments of CNF formulas with bounded clause width. Here "bounded clause width" means each clause has at most k literals for a fixed k. The algorithm is said to run in subexponential time and polynomial space. ETH is the widely used assumption that 3-SAT needs time exponential in the number of variables. Many conditional lower bounds in algorithms and optimization rest on it.

According to one report, OpenAI says the model found the disproof on 16 October and the team formalized it in Lean the next day. Both the text and the Lean proof are in the openai/math repository. Russell Impagliazzo and Mohan Paturi, two of the people who proposed ETH, are reported to be helping organize an open and independent examination.

An outside party has published a kernel-level reproduction of the Lean build. It lists the axioms used and the results of the Comparator tool. The author says it is "not an endorsement". A successful build shows the Lean proof is internally valid. Experts still have to confirm that the formal statement says what the paper claims. Commentators note that experts are still scrutinizing the proof. Until that review finishes, the disproof is a claim, not a settled result.

Tao on what Lean does and does not guarantee

On 9 October Terence Tao published a post on Lean's reliability and AI. It is commentary, not a result. It is aimed at mathematicians who will meet AI-written Lean proofs more often. He notes that as of October 2026 he knows of no complete public proof that Lean's type theory is consistent. He also discusses autoformalization, meaning AI turning mathematics into formal code. The practical point is that the check has two parts. A kernel accepts the proof, and a human confirms the formal definitions and theorem statement match the intended claim. This applies directly to the two items above.

Why it matters for optimization

If the ETH disproof survives review, the effect on optimization would be large. Many "no faster algorithm exists" arguments for combinatorial problems are built on ETH or on stronger variants. A subexponential counting algorithm for bounded-width CNF would call some of those arguments into question. That is a big "if", and the reported claim is still under independent examination. What it would change is the set of conditional lower bounds in the literature, not the practical speed of today's solvers.

The other lesson is about verification. The OpenAI repository and the independent reproduction show a workflow that optimization practitioners can reuse. A proof or algorithm claim arrives with a machine-checked artefact, and a reviewer separately audits the formal statement. The sign-error withdrawal, reported alongside the release, shows why the second step matters.

Worth watching

Sources: - OpenAI publishes 722 maths papers (The Next Web) - OpenAI community post - Quanta Magazine - Let's Data Science - Unite.ai - OpenAI - The Next Web on ETH - Technology.org - Independent Lean reproduction - Xenospectrum - Terence Tao's blog

Try ProvenDecisions free: type, photograph or say an optimization problem; every answer is checked by a solver.

What do you want to optimize?

Type the math, describe it in words, snap a photo or say it. ProvenDecisions builds the model, solves it, checks it and walks you through it.

One constraint per line, or separated by commas. Add integer: x, y or binary: z for whole-number decisions. LaTeX works too.

Try: