OpenAI Just Announced Three Scientific Breakthroughs

Each One Rewrites a Field

A Millennium Prize problem solved, a foundational computer science hypothesis disproved, and a 35-year-old conjecture cracked. All Lean-verified. All by AI.

In the year 2025, maybe we will not expect this one.

OpenAI has announced three major scientific results, one each in mathematics, computer science, and physics. All three have been formally verified in Lean, a proof assistant that checks every logical step by machine. If a proof passes Lean, the argument is valid. It is full stop, no ambiguity, no waiting for peer review to catch an error.

Each result strikes at the deepest foundations of its field. Together, they represent the most concentrated burst of AI-generated scientific discovery ever made public.

Computer science: the Exponential Time Hypothesis is false

The Exponential Time Hypothesis (ETH) has been a load-bearing pillar of theoretical computer science for decades. It concerns the computational complexity of Boolean satisfiability problems and is closely linked to the famous P versus NP question. For years, the field assumed ETH was true. Hundreds of top-tier papers used it as a starting premise, deriving computational lower bounds in graph theory, computational biology, databases, and computational geometry.

OpenAI’s model proved that ETH is false.

https://www.youtube.com/watch?v=IQJT1ScDT6c

The implications are staggering. Many problems that the field had essentially sentenced to computational death ,declaring them fundamentally too slow to solve efficiently. It may not carry that sentence at all. Core algorithms across multiple subfields could run faster than anyone believed possible. The bottom layer of theoretical computer science’s logic has been rewritten.

OpenAI obtained the natural-language disproof on October 16, 2026, and completed Lean formalization the following day. The verification team is led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego ,the very researchers who originally proposed ETH alongside Ryan Williams of MIT, who previously disproved the Strong Exponential Time Hypothesis and is among the world’s foremost experts on computational lower bounds.

At the experts’ request, OpenAI has published both the natural-language proof and the Lean code on GitHub, and has committed to funding an academic conference for the global community to examine and discuss the result.

Physics: a Millennium Prize problem, solved

In physics, quantum field theory is the bedrock of modern particle physics, but it has always lacked a rigorous mathematical foundation. The Clay Mathematics Institute identified this gap as one of seven Millennium Prize Problems, each carrying a $1 million reward: prove that Yang-Mills theory, the mathematical framework underlying the strong nuclear force, exists in a rigorous mathematical sense.

OpenAI’s model has done exactly that.

The AI produced a rigorous mathematical proof of the existence of interacting four-dimensional quantum field theories that include both gauge fields and fermions. This covers quantum chromodynamics (QCD), the core theory of the strong nuclear force, as well as a completion of the Standard Model of particle physics. As a subproblem, the proof resolves the existence of pure Yang-Mills theory .

The expert panel is led by David Gross, who won the 2004 Nobel Prize in Physics for discovering asymptotic freedom in Yang-Mills theory — the very mechanism that makes the mathematical proof possible. The panel also includes Arthur Jaffe, who co-authored the original Millennium Prize problem statement with Edward Witten and co-founded the Clay Mathematics Institute; Michael Douglas; Igor Klebanov, a pioneer in QCD strings and confinement; and Nikita Nekrasov, winner of the Dannie Heineman Prize for his work on Yang-Mills generalization methods.

The AI generated the Yang-Mills existence proof on September 9, 2026, and the QCD proof on September 18. Because the overall argument is extraordinarily large and complex, the team completed Lean formalization on October 16 before formally handing the results to the expert panel.

The physics team has chosen not to release the AI-generated raw files publicly. Instead, the panel members will write their own interpretive papers. A dedicated contact email has been established: [email protected].

In harmonic analysis, the local smoothing conjecture was proposed by Christopher Sogge in 1991. It describes how waves behave when they focus: the conjecture states that extreme focusing must be brief, and that at a typical random moment, wave focusing is far less intense than the worst-case scenario. The conjecture has profound implications across analysis, directly implying two other famous open problems: the Kakeya conjecture and the Stein restriction conjecture.

Progress on local smoothing has been painstaking. In 2019, Guth, Wang, and Zhang proved the two-dimensional case. In 2025, Hong Wang and Joshua Zahl cracked the Kakeya conjecture in three dimensions — a result that was itself a major event in mathematics.

OpenAI’s model found a new auxiliary partial differential equation and used it to prove the local smoothing conjecture in all dimensions at once, thereby resolving the Kakeya and Stein restriction conjectures completely.

The AI produced the natural-language proof on October 18 and completed Lean verification the next day. OpenAI has taken a fully open approach with this result: both the proof document and the Lean code are publicly available on GitHub for mathematicians worldwide to study and verify.

Three results, three disclosure strategies

What’s notable is that OpenAI learned from its previous controversies.The Navier-Stokes announcement in September, the disputed Erdős result in May, and the “stolen research” accusations and adopted a different public strategy for each of the three results.

For the computer science result, full transparency: natural-language proof and Lean code published on GitHub, with a funded conference for community review. This approach reflects the fact that the ETH disproof can be fully verified by machine, and the original authors of the hypothesis are themselves leading the verification.

For the physics result, expert mediation: the raw AI output is withheld, and a Nobel laureate-led panel will write the interpretive papers. This reflects the extraordinary complexity of the argument and the fact that quantum field theory proofs require deep domain interpretation that goes beyond formal verification.

For the mathematics result, open access: everything published, community invited to verify freely. This reflects the culture of mathematics, where proofs are meant to be read, challenged, and understood by the broader community.

The rules of engagement

OpenAI outlined four principles for external collaboration on these results:

Transparency. No NDAs. Participating scholars can freely discuss, share progress, and publish conclusions.

No authorship claims. OpenAI will not claim academic credit for the discoveries on behalf of its internal models.

Respect for scholarly contribution. Interpretive papers and derivative research belong entirely to the human scholars who write them, with full acknowledgment of the prior work that laid the groundwork.

Independence. Scholar participation is voluntary and unpaid. Conclusions are independent. Scholars do not represent OpenAI, and OpenAI does not represent scholars.

To support the community’s work in digesting, verifying, and extending these results, OpenAI has donated $5 million to the Simons Foundation. The funds are managed and distributed entirely by the foundation, earmarked for workshops, conferences, and research grants related to AI-generated scientific results.

What this actually means

It’s worth stepping back from the specifics to register what has just happened.

A single organization, using AI models, has in the span of weeks produced formally verified results that resolve foundational problems in three separate scientific disciplines — problems that, collectively, the world’s best human minds had worked on for decades without resolution.

The Lean verification is the critical detail. Unlike previous AI claims in mathematics and science, which relied on human referees and were sometimes disputed or retracted, a Lean-verified proof is mechanically checked at every logical step. There is no subjective judgment involved. Either the proof is valid or it isn’t. All three passed.

This does not mean the results are beyond discussion. Lean verifies logical validity, not significance or interpretation. The physics result, in particular, will require years of expert interpretation. The ETH disproof will send shockwaves through theoretical computer science, forcing the re-examination of hundreds of papers built on a premise that is now known to be false.

But the direction is unmistakable. AI is no longer assisting human scientists at the frontier. It is operating at the frontier, independently, and producing results that humans must now work to understand.

The student, as Jack Clark said last week, has become the teacher. And the teacher is not slowing down.

Sources:

• OpenAI, ETH disproof announcement and GitHub repository (October 16–17, 2026)

• OpenAI, Yang-Mills / QCD existence proof announcement (October 16, 2026; proofs generated September 9–18)

• OpenAI, local smoothing conjecture proof and GitHub repository (October 18–19, 2026)

• Clay Mathematics Institute, Millennium Prize Problems

• OpenAI, Navier-Stokes singularity proof (September 8, 2026)

• OpenAI, “Ten Advances in Mathematics and Theoretical Computer Science” (August 1, 2026)

• OpenAI, Erdős unit distance conjecture disproof (May 20, 2026)

Suggested Medium tags: Science, Mathematics, Artificial Intelligence, Physics, Technology

Suggested publications: Cantor’s Paradise (math), Towards AI, or Predict