AI Just Made Three Real Scientific Discoveries: A 35-Year Math Conjecture, a Fallen Complexity Hypothesis, and Quantum Chromodynamics

What Happened: Three Discoveries, One Announcement

On October 26, OpenAI announced that its internal AI models had produced three new scientific results — in harmonic analysis, computational complexity theory, and quantum field theory. This isn't another "AI summarizes papers" story or an AI-building-AI milestone. These are the kinds of open problems that careers are built on: a conjecture posed in 1991, a complexity hypothesis that hundreds of published results quietly depend on, and a chunk of a Clay Mathematics Institute Millennium Prize problem.

Two of the three results have already been formalized in Lean, the proof-verification language that turns a mathematical argument into something a computer can check line by line. Expert working groups — including the very people who proposed the original conjectures — are now organizing independent verification. Here's what each discovery actually is, and why researchers are calling the math result "historic."

Math: The Local Smoothing Conjecture, Proved in Every Dimension

Proposed by Christopher Sogge in 1991, the local smoothing conjecture concerns what happens to waves as they focus. Its core claim: extreme focusing must be short-lived. The amplification you see at the worst possible instant is far bigger than what you observe at a typical random moment in time.

That sounds abstract, but the conjecture sits at a crossroads of modern analysis. It implies the famous Kakeya conjecture (about how much space a needle needs to turn around in every direction) and Stein's restriction conjecture about Fourier transforms. Humans had only proved it in two spatial dimensions — Guth, Wang, and Zhang's 2019 result — with Wang and Zahl settling the 3D Kakeya case in 2025.

OpenAI says its models produced a proof of the local smoothing conjecture in every spatial dimension, via a novel auxiliary partial differential equation — which simultaneously resolves the Kakeya and Stein restriction conjectures in all dimensions. The AI output arrived October 18 and was formalized in Lean the next day. Alex Iosevich, a University of Rochester harmonic analyst (and Sogge's former PhD student), called it a "historic development."

Computer Science: The Exponential Time Hypothesis Is False

The Exponential Time Hypothesis (ETH), proposed in the late 1990s by Impagliazzo, Paturi, and Zane, conjectures that there is no subexponential-time algorithm for Boolean satisfiability — the canonical hard problem sitting next to P vs. NP. It has been widely believed to be true, and it quietly underpins hundreds of conditional hardness results: proofs that start with "assuming ETH, problem X cannot be solved faster than…" across graph theory, computational biology, databases, and geometry.

OpenAI's system found a deterministic algorithm that exactly counts satisfying assignments of fixed-width CNF formulas in subexponential time and polynomial space. The 3-SAT case of that algorithm refutes the ETH outright. The disproof was found October 16 and verified in Lean on October 17. MIT's Ryan Williams — one of the field's leading figures — told Al Jazeera that a confirmed refutation would be "a major development," noting that "a great deal of theoretical computer science was built on it being true," while cautioning that the working group is still verifying the proof.

If it holds, the consequences ripple outward: every theorem conditioned on ETH needs reexamination, and the door opens to genuinely faster algorithms for problems everyone had written off as intractable.

Physics: Quantum Chromodynamics Gets an Existence Proof

The third result is in constructive quantum field theory. OpenAI's models reportedly produced a mathematically rigorous existence proof of interacting quantum field theories with gauge fields and fermions in four dimensions — including Quantum Chromodynamics (QCD), the theory of quarks, gluons, and the strong force, and a unified completion of the Standard Model. As a subproblem, it includes a proof of the existence of pure Yang–Mills theory, part of the Clay Millennium Prize problem that has stood open (with a $1 million bounty) since 2000.

QCD has been spectacularly successful as physics — it is the theory, confirmed by decades of collider data — but nobody had ever shown it exists as a mathematically well-defined object. The AI produced a Yang–Mills existence proof on September 9 and a QCD proof on September 18; OpenAI delayed release until the Lean formalization of Yang–Mills completed on October 16.

The verification group here is as heavyweight as it gets: David Gross (2004 Nobel laureate for asymptotic freedom, the key mechanism behind these existence results), along with Michael Douglas, Arthur Jaffe (who co-formulated the Millennium problem and co-founded the Clay Institute), Igor Klebanov, and Nikita Nekrasov. At their request, the raw physics artifacts are being held back while the group writes the work up itself.

The New Playbook: Shepherds, No Credit, Lean Formalization

Almost as notable as the results is how OpenAI released them. Describing the AI manuscripts as "raw research data," the company asked external experts to act as shepherds and tried three different release strategies: publish the math output raw; publish the CS output alongside an organized external collaboration; and defer entirely to the physicists on the physics artifacts. Four principles govern the process:

• Transparency — no NDAs; participants are free to discuss and publish.
• No academic credit — OpenAI seeks none for discoveries made by its internal models.
• Recognition — the human groups get credit for the understanding and writeups they produce.
• Independence — participation is voluntary, unpaid, and the groups don't speak for OpenAI.

The company is also backing a $5 million contribution to the Simons Foundation to fund workshops and programs helping researchers unpack AI-produced results. The math and CS writeups, plus the Lean proofs, are public on OpenAI's GitHub repository.

What This Means for How You Use AI Tools

You're probably not proving conjectures this week, but this shift — from AI that drafts text to AI that produces verifiable, checkable knowledge — changes what to expect from the tools you pick:

• Verification is the product now. The Lean formalizations are what make these claims credible rather than another hallucination story. Favor tools that show their work: citations, traceable steps, machine-checkable outputs. That's exactly what serious research assistants like Elicit and Consensus are built around.
• Deep-research workflows are table stakes. Multi-hour autonomous investigation used to be a frontier demo; it just produced Millennium-adjacent results. ChatGPT, Claude, and Perplexity Pro all now ship long-horizon research modes — pick the ones that cite sources you can audit.
• Formal verification is going mainstream. If AI provers can be pointed at math, they can be pointed at code correctness too — expect proof-assisted guarantees to start appearing in developer tooling like Claude Code and Codex.
• Specialists still close the loop. Every one of these results is being vetted by human experts before it counts. The pattern for the foreseeable future is AI proposes, humans verify — not AI alone.

Frequently Asked Questions