OpenAI Says It Disproved the Exponential Time Hypothesis and Wants Academics to Check the Work

News Summary

On October 26, 2026 (Eastern Time), OpenAI announced three new research results produced by its internal AI models, one each in mathematics, computer science and theoretical physics, together with a new plan for how outside academics will examine and explain them. The headline claim is that the Exponential Time Hypothesis, a long-standing assumption about how fast hard computational problems can be solved, is false. The other two cover a proof related to the Kakeya conjecture in wave analysis and an existence result for certain quantum field theories. All three are claims that are still under independent review.

The Three Results

The first result is in mathematics. OpenAI released a manuscript proving Sogge's local smoothing conjecture for waves in every spatial dimension. Gizmodo reports that the work also resolves the Kakeya conjecture, a problem from the study of waves and harmonics. The manuscript was published directly so that researchers can examine it and write their own explanations.

The second result is in computer science. OpenAI says its system produced a proof that the Exponential Time Hypothesis (ETH) is false. ETH is widely believed to be true and is used as a benchmark for estimating how long hard problems such as 3-SAT should take to solve. According to Dataconomy, it underpins hundreds of papers in areas such as databases, computational biology and geometry. If the disproof holds, faster algorithms for some of those problems could become possible in theory. The scientists who originally proposed ETH are among the experts asked to examine the model-generated work.

The third result is in physics. OpenAI describes a proof for certain interacting quantum field theories with gauge fields and fermions in four dimensions, including Quantum Chromodynamics (QCD), which describes the strong force, and a unified completion of the Standard Model. A separate expert group is reviewing the physics material and had not published anything about it at the time of reporting. That group decided to hold back the AI-produced artifacts and write up the work itself.

A New Way to Work With Academics

Alongside the results, OpenAI described how it wants to publish them. It says it does not seek academic credit for discoveries made by its internal models. It describes the model-generated texts as "best understood as a form of raw research data" that math and science communities can turn into scientific understanding.

External experts, often scholars whose earlier work the models built on, act as "shepherds." Each shepherd assembles a group to review the output and produce explanations. Group members receive credit for the writeups and materials they create, and they are responsible for making sure the results credit the researchers, techniques and ideas the models drew on. Some reports say the computer science group is led by Russell Impagliazzo and Ramamohan Paturi, and that OpenAI has committed to publishing raw data and funding a conference. These details came from secondary reports and we could not confirm them against OpenAI's own page.

OpenAI also acknowledged a limit of the approach: choosing the first shepherds shapes who gets involved early and creates an information asymmetry. It invited feedback on broadening participation. Some of the raw material is on GitHub, but not all of it.

Context and Caveats

The announcement follows an earlier release at the start of October 2026, when OpenAI posted hundreds of pages of manuscripts on GitHub. Reports give the count as 377 math results (Gizmodo) or more than 700 pages covering 372 major problems (Dataconomy). Gizmodo noted that the accompanying blog post read as somewhat defensive, and that an advisory group OpenAI had worked with had criticized a private company using unreleased internal models to reach such breakthroughs.

Verification is still an open question. Gizmodo reports that math results were formalized in Lean, a proof assistant, but that human readers examining OpenAI's Lean-verified solution to a Navier-Stokes problem raised concerns that Lean had checked a weaker statement than the one OpenAI published. Reports about whether the ETH disproof has been formalized in Lean differ between sources. The underlying models are not publicly available, so outside researchers cannot reproduce how the results were generated, and none of the three claims has yet been independently confirmed.

Why It Matters for Science Education

Whether or not every claim survives review, the episode shows how research may work as AI systems begin to produce candidate proofs faster than people can read them. Checking, explaining and teaching the results becomes as important as finding them. For students and teachers, the practical lesson is that machine-generated proofs still need human understanding, formal tools like Lean, and open discussion before they count as established knowledge.

Sources

Reporting is drawn from Gizmodo (October 26, 2026, 8:13 p.m. Eastern Time), Dataconomy (October 27, 2026), and summaries of OpenAI's post "Working with academic communities to build scientific understanding." The Verge article named as a starting reference could not be retrieved.