Connect with us

NEWS

OpenAI Hands Its ETH Claim to Unpaid Academic Shepherds

OpenAI posted Lean-checked claims on ETH, Kakeya, and Yang-Mills, named unpaid shepherds, and walked away from credit as a second lab posted a twin ETH disproof.

Published

on

OpenAI said on October 26 that an internal model had refuted the Exponential Time Hypothesis, a 1999 conjecture that 3-SAT cannot be solved in subexponential time. The company posted a write-up and a Lean proof, named unpaid academic shepherds to explain the work, and said it does not want academic credit.

The same post added a proof of Sogge’s local smoothing conjecture in every dimension and a rigorous existence argument for certain four-dimensional quantum field theories, including QCD. None of the three claims has been certified by the people asked to check them. Hours later, Anthropic posted a second ETH disproof that its own model found on the same October 16 date.

OpenAI Calls the Proofs Raw Data

The company framed the drop as an experiment in how a lab should hand machine output to working scientists. It called the manuscripts as raw research data and asked outside experts to act as shepherds, organizing groups that would try to understand the results and explain them. Those shepherds then chose three different release paths, one per field.

Math went straight to GitHub, with researchers left to organize themselves. Computer science went public too, but with a standing group led by the scientists who first proposed ETH. Physics stayed closed: a group led by David Gross chose to hold the AI files and write the work up themselves. OpenAI said it does not yet know which path helps, and it admitted that picking the first shepherds “creates asymmetry of information.”

THE THREE CLAIMS OPENAI POSTED

Field Claim AI output Lean check Public files Shepherds
Mathematics Local smoothing in every dimension, which would also settle Kakeya and Stein restriction October 18, 2026 October 19, 2026 Write-up and Lean proof on GitHub No standing group; one expert advised a public dump
Computer science The Exponential Time Hypothesis is false October 16, 2026 October 17, 2026 Write-up and Lean proof on GitHub Russell Impagliazzo, Ramamohan Paturi, Ryan Williams
Physics Existence of certain 4D gauge theories, including QCD, a Standard Model completion, and pure Yang-Mills Yang-Mills September 9; QCD September 18, 2026 Yang-Mills October 16, 2026; QCD not stated Held by the group David Gross, Michael Douglas, Arthur Jaffe, Igor Klebanov, Nikita Nekrasov

Lean is a language in which a computer checks each step of a proof. OpenAI used it on the math and computer science claims, and on Yang-Mills existence, then still asked humans to decide what the checked statements mean for their fields. A machine that accepts a Lean file has not certified that the file is the conjecture working scientists had in mind.

What a Fallen ETH Would Change

Russell Impagliazzo and Ramamohan Paturi of UC San Diego proposed ETH in 1999, later with Francis Zane. It says there is some constant c such that no deterministic algorithm decides 3-SAT in time 2^{c n}. Hundreds of papers have used that assumption to derive conditional hardness results in graph theory, computational biology, databases, and computational geometry.

OpenAI says its system proved the hypothesis false, a result it then checked in Lean. MIT professor Ryan Williams, posting a link to OpenAI’s ETH disproof paper, described the claim as placing 3-SAT on n variables in about exp(n/(log n)^{1/2+o(1)}) time. That bound is subexponential, which is all a refutation needs. It is still far from polynomial, so P versus NP stays open either way. ETH, if true, would imply P is not NP; the converse is not known, and a subexponential SAT algorithm would not settle it.

OpenAI itself said that turning the argument into faster algorithms people would run “will require significant further research.” The practical reading is blunt. SAT solvers already crush many easy instances, and a bound that only pulls ahead at enormous n does not rewrite encryption or chip design. What it would rewrite, if it holds, is a large body of “assuming ETH…” lower bounds. The reductions and techniques in those papers would still be there. Some of their conclusions would need new foundations, or weaker conjectures such as Orthogonal Vectors, to stand on.

The acronym did other damage on the way out. ETH is also the ticker people use for ether, so a complexity claim moved through finance feeds that were not looking for 3-SAT. That confusion is a sideshow. The load-bearing question is whether the Lean statement is the Exponential Time Hypothesis as Impagliazzo and Paturi defined it, on the machine model theorists actually use.

Unpaid Reviewers With No Power to Certify

The group asked to examine the computer science claim is led by Impagliazzo, Paturi, and Williams. OpenAI notes that Williams has previously refuted strong versions of ETH and has also used the weak form for hardness results. In a letter to colleagues, the three said their names may appear with the release and then drew a hard line under what that appearance means.

Our involvement should not be taken as certification of the claim. A result of this significance requires careful, independent scrutiny and an understanding that others can reconstruct, question, and build upon.

Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams, letter to the theoretical computer science community

Williams said the same in public, in four posts that also set the terms of the job. The three signed no NDAs and will receive no pay. They asked OpenAI to release all raw ETH data, including PDFs and Lean source, and to fund a conference for the field affected by the work. He said the company agreed to both.

OpenAI listed four rules for these outside groups, and they read as much like a liability shield as a gift.

THE FOUR RULES FOR SHEPHERDS

  • Transparency: No nondisclosure agreements, and participants may discuss the work, share drafts, and publish their own conclusions.
  • No academic credit: OpenAI does not seek authorship for discoveries produced by its internal models, and inviting a shepherd does not mean that person made the find.
  • Recognition: The group should get credit for the understanding it develops and the write-ups it authors, and should credit earlier ideas the proofs use.
  • Independence: Work is voluntary and unpaid; members do not speak for OpenAI, and OpenAI does not speak for them.

The letter goes further than those rules. The three asked how the community should organize an open exam so that no company and no selected group, including themselves, owns it. They asked how credit should treat the unglamorous work of checking, fixing, simplifying, and explaining. They asked how to support students as research directions move. They floated a virtual meeting before the end of 2026. If the claim is correct, they wrote, “it will change an important part of theoretical computer science.”

That is the bind the announcement built. OpenAI keeps the headline and refuses the authorship. University scientists, some of whom spent careers on the hypothesis now on the table, inherit the months of checking and the career risk if the proof is wrong, or merely if the Lean statement is not the theorem the field thought it was.

A Second Lab Posted a Twin Proof Overnight

On October 27, Anthropic added a folder to its public formal-math repository titled “A 2^{o(n)} Time Algorithm for 3-SAT,” credited to Claude (Anthropic). The paper is dated October 26. The GitHub commit landed at 03:07 UTC on October 27, hours after OpenAI’s own files. Anthropic’s paper says the core algorithm was found on October 16 by an internal Claude research model with 14 million output tokens and no human input, after which Anthropic researchers worked with Claude on the presentation. It notes, in one line, that an OpenAI model claims a similar result independently and concurrently.

The method is not a copy. Accounts of the two papers describe OpenAI’s argument in terms of linear summaries and sparse polynomial systems, and Anthropic’s in terms of tensor-network contraction and rank bounds. The leading term in the exponent is in the same family. Anthropic’s README states a deterministic algorithm that counts satisfying assignments of k-CNF formulas in time 2^{O(n log log n / √log n)}, and it says a 3-SAT decision algorithm in 2^{o(n)} time refutes ETH. It also says, in a sentence the field needed, that the result is purely asymptotic and says nothing about solving formulas in practice.

The Lean 4 project, which Anthropic says was written by AI models, is pinned to Lean v4.33.1 and the matching Mathlib. A full build took about 8 minutes on 24 virtual CPUs, with the largest process using about 7 GB of memory. The README says the proved theorems depend only on Lean’s three standard axioms: propext, Classical.choice, and Quot.sound. Continuous integration also re-checks the proofs with nanoda, a second, independently written kernel, and compares the theorem statements with a short Challenge file. Even that stack still leaves a human job that no kernel does: confirming that MainResults.lean says what the PDF claims, and that both say what ETH means in the literature.

Two labs, two write-ups, one find date, and still no community verdict. The twin does not make OpenAI’s file correct. It does make the “who saw it first” fight smaller than the “does it hold” fight, and it shows that a closed model at more than one company was pointed at the same conjecture in the same window.

Kakeya in Every Dimension Lands on GitHub

The mathematics claim is the one OpenAI treated most like its earlier October 6 dump of AI-written math papers. Christopher Sogge proposed the local smoothing conjecture in 1991. It is a statement about waves: the most extreme focusing must be short-lived, and the focusing at a typical time is milder than the worst case. OpenAI says that because the phenomenon is so basic, the conjecture implies the Kakeya conjecture and Stein’s restriction conjecture.

Human proofs had reached only pieces. Larry Guth, Hong Wang, and Ruixiang Zhang proved local smoothing in two spatial dimensions in 2019. Wang and Joshua Zahl proved the Kakeya conjecture in three dimensions in 2025. OpenAI says its models proved local smoothing in every dimension, using a new auxiliary partial differential equation, and that this also resolves Kakeya and Stein restriction in all dimensions. The AI write-up is dated October 18, the Lean proof October 19. Both are on GitHub. OpenAI says it spoke with one outside expert, who recommended a public release rather than a standing shepherd group.

That is the cleanest of the three paths, and it is still not a journal acceptance. A GitHub PDF plus a Lean file is an invitation to argue, which is what the October 6 math collection already forced the field to do. The computer science group at least has named organizers and a letter asking for help. The math claim has a repo.

FROM SEPTEMBER TO THE PUBLIC FILES

  1. September 9, 2026: OpenAI’s models produce a natural-language proof of Yang-Mills existence.
  2. September 18, 2026: The models produce a proof for QCD.
  3. October 16, 2026: OpenAI finishes a Lean check of Yang-Mills existence. The same day, its model finds a natural-language ETH disproof, and Anthropic says Claude finds the core of its own 3-SAT algorithm.
  4. October 17, 2026: OpenAI formalizes the ETH claim in Lean.
  5. October 18 and 19, 2026: Local smoothing output, then its Lean proof.
  6. October 25, 2026: The Advisory Group on Mathematics and Artificial Intelligence publishes a note against labs designating mathematicians to receive AI output first.
  7. October 26, 2026: OpenAI announces all three results. AGMAI, the same day, calls for immediate public release of the Yang-Mills and QCD files.
  8. October 27, 2026: Anthropic posts its ETH paper and Lean 4 project.

The physics dates sit on that list for a reason. OpenAI says it waited to announce until the Yang-Mills Lean work finished, then handed release control to Gross’s group, which chose to keep the artifacts in house.

Why the Yang-Mills Files Are Still Private

The physics claim is the one that reaches for a Clay problem, and the one nobody outside a named circle can read. OpenAI says its models produced a mathematically rigorous existence proof of certain interacting four-dimensional quantum field theories with gauge fields and fermions, including QCD and a unified completion of the Standard Model, building on Tadeusz Balaban’s constructive approach. As a subproblem, it says, the work includes a proof that pure Yang-Mills theory exists, which is part of the Yang-Mills Millennium Prize problem.

That prize problem, as Clay has long stated it, is not existence alone. It also asks for a mass gap. OpenAI’s post claims the existence piece, not the gap, so the claim as written does not trigger the prize. The company is also careful about what the Standard Model part covers: the existence proof, it says, pertains to the simplest grand unified theory that includes the Standard Model’s gauge fields and fermions, because the Standard Model itself cannot be pushed to arbitrarily short distances.

Gross will lead the group. His 1973 work on asymptotic freedom in Yang-Mills theories helped establish QCD as the theory of the strong force and earned him a share of the 2004 Nobel Prize. The other named members are Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov. Jaffe formulated the Clay Yang-Mills problem with Edward Witten and helped found the Clay Mathematics Institute. OpenAI says the group asked that all contact go through a single email account, and that it will write the work up itself rather than post the model’s files.

AGMAI, an independent nine-member panel hosted at the Institute for Advanced Study, did not wait. On October 25 it had already argued that labs should not designate particular mathematicians to receive AI output, because that choice gives a head start, denies the field agency, reflects the companies’ networks, and lets famous names vouch for a release the names did not produce. On October 26, the day of OpenAI’s post, it called on the company to make the natural-language and Lean outputs for Yang-Mills, QCD, and the Standard Model completion public without delay. OpenAI had said it delayed until Lean was done. “By their own account this is now complete,” AGMAI wrote, “so there is no scientific reason for delaying the release any further.”

The $5 million OpenAI recently sent to the Simons Foundation, for workshops and programs on AI-produced results, does not answer that demand. Money for later meetings is not the same as posting the files the Gross group is holding. OpenAI’s own four principles say participants will not be asked to sign NDAs and will be free to discuss the work. The physics artifacts are still not on GitHub.

WHAT WE KNOW

  • Public math and CS files: OpenAI posted natural-language write-ups and Lean proofs for local smoothing and for the ETH claim.
  • Named, unpaid examiners: Impagliazzo, Paturi, and Williams are organizing an open look at ETH and have said they do not certify it.
  • A second ETH paper: Anthropic posted a different Lean-checked 3-SAT algorithm, also traced to October 16, and warned it is asymptotic only.
  • Prize scope: The physics write-up OpenAI describes covers existence, including pure Yang-Mills, not the mass gap.

WHAT IS UNCONFIRMED

  • Whether ETH is false: No independent reconstruction has been announced, and the shepherds have said their names are not a stamp.
  • Whether the Lean statements match the conjectures: A kernel can accept a proof of the wrong theorem.
  • The physics artifacts: The natural-language proofs and the Yang-Mills Lean file remain with Gross’s group.
  • Practical algorithms: Both labs describe bounds that do not, on their face, yield faster SAT solvers anyone would run.

The three results may yet hold. If they do, the first people to make them usable will be the unpaid groups now reading machine text, not the company that posted the headline. The files that are public are on GitHub. The files that are not sit with a closed physics group, on the same day an independent math panel asked for them to be opened.

Harry is the editor of IAQABA, an independent publication he owns and runs. A decade in journalism, beginning as a reporter and now as the editor of his own titles, has left him with a clear test for what deserves a story: it has to change what a reader knows or decides, and it has to rest on something he can point to. That rules out recycled press releases, forecasts with no data behind them and rumours that no document supports. It leaves room for a great deal, and the site covers news, business, science and technology alongside sports, entertainment and lifestyle, with travel, auto and gaming given the same standard rather than lighter treatment. Sources are primary wherever possible: the regulator's filing, the company's own statement, the transcript, the dataset, or the product on Harry's desk. Figures are checked before they are published and rechecked if a reader questions them. Mistakes are corrected under a published policy. Readers across the world can reach him directly at support@iaqaba.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending