AI Claimed Three Scientific Breakthroughs in One Day. Proof Is Now the Bottleneck

Guide

Core argument

On October 26, one lab claimed three scientific results at once — and the real story is not whether machines can discover, but whether institutions can absorb machine-made science. The bottleneck has moved from finding proofs to trusting them.

Counter-intuitive point

A machine-checked proof is not the same as a settled theorem. See the 'gap Lean cannot close' branch for why formal verification can pass while the headline still fails.

What to do next

Before sharing the next 'AI proved X' headline, run the checklist in the final branch — four checks that separate a verified result from a press release.

Connections

This piece extends two earlier Library entries: Gödel's Incompleteness Theorems: The 24-Year-Old Who Proved Every System Has Blind Spots and The 24-Year-Old Who Proved Computers Can't Do Everything — machines are now working inside the very limits those two pieces mapped.

Outline

• AI Claimed Three Scientific Breakthroughs in One Day. Proof Is Now the Bottleneck
  • The announcement: October 26, 2026, three claims at once
    • OpenAI said its internal models produced all three results, presented as unreviewed research preprints, not peer-reviewed papers
    • The company says it is not seeking credit for the results; outside academic groups will write them up
    • A timeline inside the announcement: physics outputs dated September 9 and 18, the complexity result October 16, the math proof October 18
    • On October 6, OpenAI had already released 722 mathematics manuscripts in an earlier research dump
    • Why one day matters: three fields, three different verification paths, one shared question — who checks machine-made science?
  • Claim one: waves, needles, and a 35-year-old conjecture
    • Sogge's local smoothing conjecture, proposed 1991: how waves spread out and smooth over time
    • The payoff: the proof is said to resolve the Kakeya conjecture and the Stein restriction conjecture in all dimensions
    • The Kakeya needle problem, posed 1917: a set containing a unit segment in every direction can still have zero area
    • Human progress before the machine: Guth, Wang and Zhang proved local smoothing in 2 dimensions in 2019; Wang and Zahl proved Kakeya in 3 dimensions in 2025
    • The machine's route: a novel auxiliary partial differential equation, output October 18, Lean formalization October 19
    • Status: natural-language manuscript and Lean materials public on GitHub
  • Claim two: the hypothesis that hundreds of papers stood on
    • The Exponential Time Hypothesis (ETH), proposed by Impagliazzo, Paturi and Zane: 3-SAT cannot be solved in subexponential time
    • Why it mattered: used as a hardness assumption in hundreds of papers across graph theory, computational biology, databases and geometry
    • The claim: a deterministic algorithm refutes it; output October 16, Lean formalization October 17
    • The reviewers: a group led by the hypothesis's own proposers Impagliazzo and Paturi, with MIT professor Ryan Williams
    • Their involvement is not an endorsement, the reviewers say — examination first, verdict later
    • A second lab is reported to have reached the same conclusion by a different route — a cross-check to watch, not yet a confirmation
    • If it falls, conditional hardness results built on ETH must be re-examined, not automatically discarded
  • Claim three: the physics result nobody can read yet
    • A mathematically rigorous existence proof for four-dimensional quantum field theories, including quantum chromodynamics (QCD)
    • As a subproblem: existence of pure Yang-Mills theory, part of a Clay Millennium Prize problem (7 problems, 1 million dollars each)
    • The route: building on Balaban's approach to constructive quantum field theory; Yang-Mills output September 9, QCD September 18
    • Lean formalization of the Yang-Mills proof finished October 16; no formalization mentioned for the QCD proof
    • The materials are withheld while a group led by Nobel laureate David Gross prepares its own explanation first
    • Pushback the same day: the Institute for Advanced Study's Advisory Group on Mathematics and AI publicly called for immediate release
    • Counter-intuitive: the most spectacular claim is the least checkable — the physics proof exists, for now, as an announcement about a proof
  • How machines learned to prove: a short history
    • 1976: Appel and Haken's computer-assisted proof of the Four Color Theorem — mathematicians argued for years over whether a proof no human can read counts
    • 1998-2014: Hales's proof of the Kepler conjecture, then the Flyspeck project formally verifying it in a proof assistant
    • 2023: DeepMind's FunSearch finds a new construction for the cap set problem — machine output, human-verified
    • 2024: AlphaGeometry and AlphaProof reach olympiad-medal level on geometry and competition problems
    • The infrastructure underneath: Lean and its mathlib library, where proofs are code a computer can check line by line
    • The pattern: each step moved the argument from 'can a machine help?' to 'can a machine originate?'
  • The shepherd model: a new way to release a result
    • Instead of one paper and one journal, each claim gets its own verification path chosen by the academics it affects most
    • Math and computer science: manuscripts plus Lean code public, review groups named
    • Physics: a closed review group writes first, materials later
    • What is new: the discoverer steps back — OpenAI hands authorship of the write-up to outside groups
    • What is old: preprints before peer review, the same route that carried the 2023-2024 AI results
    • The risk: different rules per field mean different standards of evidence under one headline
  • The gap Lean cannot close
    • Lean checks that a formal statement follows from axioms — it cannot check that the formal statement is the theorem in the headline
    • Independent tracking of the earlier October 6 release already distinguishes machine-checked statements from full validation of advertised claims
    • The ETH case shows the gap in public: Lean materials describe subexponential-time results for fixed-width Boolean formulas, while reviewers judge whether that refutes ETH
    • A formal proof can be correct and still be a weaker claim than the press release — or a stronger one than humans yet understand
    • Counter-intuitive: the bottleneck is not computing power but translation — turning a machine's argument into a human's reason to believe
    • What would close it: an independent group reproducing the key lemma, in its own words, from the Lean code
  • How to read the next 'AI proved X' headline
    • Checklist — four questions before you share
    • Is the Lean code or formal proof public, or only the announcement?
    • Does the formal statement match the headline claim, word for word?
    • Who is shepherding verification — the claimants, the skeptics, or the original proposers?
    • Has any independent group reproduced the key step, not just re-run the code?
    • Decision rule: public code plus independent reproduction equals a result; an announcement plus a review group equals a claim to watch
    • What to watch next: the ETH reviewers' report, the Gross group's physics write-up, and whether journals change their rules for machine authors

Node details

AI Claimed Three Scientific Breakthroughs in One Day. Proof Is Now the Bottleneck

The announcement: October 26, 2026, three claims at once

OpenAI said its internal models produced all three results, presented as unreviewed research preprints, not peer-reviewed papers

The company says it is not seeking credit for the results; outside academic groups will write them up

A timeline inside the announcement: physics outputs dated September 9 and 18, the complexity result October 16, the math proof October 18

On October 6, OpenAI had already released 722 mathematics manuscripts in an earlier research dump

Why one day matters: three fields, three different verification paths, one shared question — who checks machine-made science?

Claim one: waves, needles, and a 35-year-old conjecture

Sogge's local smoothing conjecture, proposed 1991: how waves spread out and smooth over time

The payoff: the proof is said to resolve the Kakeya conjecture and the Stein restriction conjecture in all dimensions

The Kakeya needle problem, posed 1917: a set containing a unit segment in every direction can still have zero area

Human progress before the machine: Guth, Wang and Zhang proved local smoothing in 2 dimensions in 2019; Wang and Zahl proved Kakeya in 3 dimensions in 2025

The machine's route: a novel auxiliary partial differential equation, output October 18, Lean formalization October 19

Status: natural-language manuscript and Lean materials public on GitHub

Claim two: the hypothesis that hundreds of papers stood on

The Exponential Time Hypothesis (ETH), proposed by Impagliazzo, Paturi and Zane: 3-SAT cannot be solved in subexponential time

Why it mattered: used as a hardness assumption in hundreds of papers across graph theory, computational biology, databases and geometry

The claim: a deterministic algorithm refutes it; output October 16, Lean formalization October 17

The reviewers: a group led by the hypothesis's own proposers Impagliazzo and Paturi, with MIT professor Ryan Williams

Their involvement is not an endorsement, the reviewers say — examination first, verdict later

A second lab is reported to have reached the same conclusion by a different route — a cross-check to watch, not yet a confirmation

If it falls, conditional hardness results built on ETH must be re-examined, not automatically discarded

Claim three: the physics result nobody can read yet

A mathematically rigorous existence proof for four-dimensional quantum field theories, including quantum chromodynamics (QCD)

As a subproblem: existence of pure Yang-Mills theory, part of a Clay Millennium Prize problem (7 problems, 1 million dollars each)

The route: building on Balaban's approach to constructive quantum field theory; Yang-Mills output September 9, QCD September 18

Lean formalization of the Yang-Mills proof finished October 16; no formalization mentioned for the QCD proof

The materials are withheld while a group led by Nobel laureate David Gross prepares its own explanation first

Pushback the same day: the Institute for Advanced Study's Advisory Group on Mathematics and AI publicly called for immediate release

Counter-intuitive: the most spectacular claim is the least checkable — the physics proof exists, for now, as an announcement about a proof

How machines learned to prove: a short history

1976: Appel and Haken's computer-assisted proof of the Four Color Theorem — mathematicians argued for years over whether a proof no human can read counts

1998-2014: Hales's proof of the Kepler conjecture, then the Flyspeck project formally verifying it in a proof assistant

2023: DeepMind's FunSearch finds a new construction for the cap set problem — machine output, human-verified

2024: AlphaGeometry and AlphaProof reach olympiad-medal level on geometry and competition problems

The infrastructure underneath: Lean and its mathlib library, where proofs are code a computer can check line by line

The pattern: each step moved the argument from 'can a machine help?' to 'can a machine originate?'