OpenAI claims three AI advances in math, computing and physics

By Rebecca Stone

OpenAI said on Oct. 26 that internal AI models had produced three claimed results on long-standing problems in mathematics, computer science and physics, with writeups and Lean formalizations released for the mathematics and computer-science work.

The statement came nearly seven weeks after OpenAI announced an AI-generated proposed solution to the Navier–Stokes problem and three weeks after it released hundreds of mathematical manuscripts. The earlier Navier–Stokes offering was a proposal rather than one of the three new results.

The sequence documents repeated AI-assisted research outputs, but the announcement gives no model run counts, model identities or division of labor between machine generation and human checking. For research groups assessing reproducibility, those omissions limit any comparison of effort, compute use or scientific workflow.

What did OpenAI claim?

• Computer science: Its models reportedly refuted the Exponential Time Hypothesis (ETH), a decades-old conjecture about how quickly computers can solve certain logic problems. OpenAI says the conjecture underpins hundreds of studies of computational limits, so a valid refutation could alter assumptions and methods across that research field.
• Mathematics: The models reportedly proved the local smoothing conjecture in every spatial dimension. OpenAI describes this as resolving a long-standing question about how waves concentrate and says the result carries implications for the Kakeya and Stein restriction conjectures.
• Physics: OpenAI claims a rigorous mathematical construction of interacting quantum field theories in four dimensions, including quantum chromodynamics, which describes the strong force binding quarks inside protons and neutrons.
• Yang–Mills scope: The physics work includes an existence proof for pure Yang–Mills theory. OpenAI calls that part of the Clay Mathematics Institute’s Yang–Mills Millennium Prize problem. The full Clay problem also requires a mass gap; this announcement covers existence alone.

Why does the computer-science claim matter?

OpenAI co-founder and president Greg Brockman wrote that AI is “starting to help us improve our understanding of the universe.”

Russell Impagliazzo and Ramamohan Paturi of UC San Diego and Ryan Williams of MIT wrote to their research community: “If the claim is correct, it will change an important part of theoretical computer science.”

They said the Exponential Time Hypothesis has connected decades of research on computational difficulty and that techniques developed around it retain value even if the interpretation changes. They also said their involvement did not certify OpenAI’s claim. The qualification matters because a purported refutation would require researchers to reassess where their arguments rely on the conjecture, even if the underlying algorithms remain useful.

What can the research community check?

OpenAI has published writeups for the mathematics and computer-science results and formalizations in Lean, software for checking formal mathematical proofs. A Lean file gives reviewers a machine-checkable audit trail, comparable to a software test suite.

The announcement does not identify the model versions, prompting procedure, compute budget, training data or funding for the work. These are proof claims, so experimental sample sizes and confidence intervals do not apply. Review instead depends on the complete logical chain and its cited results. The immediate workflow centers on proof checking and specialist review; the announcement reports no independent validation or experimental replication.

The independent Advisory Group on Mathematics and Artificial Intelligence called for immediate release of the physics proofs, saying everyone should have the same access to the work. OpenAI says the external physics group chose to withhold the AI-produced materials while preparing its own account. The announcement therefore supplies no physics derivation for community checking.

What did Anthropic add?

Hours after OpenAI’s announcement, Anthropic posted a separate claimed refutation of the Exponential Time Hypothesis, credited to “Claude (Anthropic).” The paper says an internal Claude research model found the core algorithm without human input on Oct. 16, the same discovery date OpenAI reports. Anthropic researchers subsequently helped improve its presentation, and the release includes a Lean formalization.

Anthropic cautioned that the theoretical result does not establish a practical speedup for solving real-world problems. That caveat separates algorithmic possibility from deployable performance, while the matching Oct. 16 date provides a cross-company point of comparison without amounting to independent certification.

What should R&D managers do with the claims?

For research portfolios, treat the three outputs as potentially consequential but unverified inputs until external proof checking is complete. The review burden differs by claim: two have published derivations and Lean artifacts; the physics claim lacks released materials; Anthropic supplies another formalization but says no practical speedup follows.

Neither OpenAI nor Anthropic discloses, in the announcement described here, the compute cost or funding behind the model runs. Managers therefore cannot benchmark resource intensity or reproduce the workflow from these releases.

The next portfolio decision point is access to the physics derivation and independent scrutiny of each formalization. Until then, the announcements establish claims and reviewable artifacts, not completed scientific wins.