AI/Nov 17 edition/Video analysis
ResearchModels

ResearchPapers, evidence & method

Nobel laureate David Gross's committee affirms OpenAI models' Lean proof that 4D Yang–Mills theory exists, leaving the mass gap unsolved

A committee led by Nobel laureate David Gross says a Lean 4 proof generated by OpenAI's models establishes that pure four-dimensional Yang–Mills theory exists. The harder mass-gap half of the Clay Millennium problem remains open, and the sprawling code is hard to read.

From the Two Minute Papers video of November 17, 2026 and cited sources: The Gross committee says the machine-checked proof settles the existence half of the Clay Yang–Mills problem under the standard problem statements. [2] [3] [6] [8]

SOURCES[2] Scientific American: Scientists confirm OpenAI's… scientificamerican.com[3] GitHub: yang-mills-group/OpenAI-YM-20261028 github.com[6] Pasquale Pillitteri: Committee confirms OpenAI's… pasqualepillitteri.it

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
The Gross committee says the machine-checked proof settles the existence half of the Clay Yang–Mills problem under the standard problem statements. [2] [3] [6] [8]
Original source
Two Minute Papers video, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
November 17, 2026 video published
Evidence
8 sources across 7 sites; transcript sourced
45/100RisingHow it’s measured

YouTube video · 15.5K views · 864 likes

Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

Illustration from Two Minute Papers: Nobel laureate David Gross's committee affirms OpenAI models' Lean proof that 4D Yang–Mills theory exists, leaving the mass gap unsolved
Image: Two Minute Papers — Original article ↗
THE CORE IDEAS3 TAKEAWAYS
01

The Gross committee says the machine-checked proof settles the existence half of the Clay Yang–Mills problem under the standard problem statements. The main theorem builds continuum limits that satisfy the Wightman axioms and shows the theory is not free. The mass gap remains unproven. [2] [3] [6] [8]

02

The construction extends Bałaban's lattice renormalization-group program. It tracks several grid scales at once so that fine-scale quantum fluctuations stay controlled as the lattice is refined. This ultraviolet machinery makes up roughly 70% of the code. [1] [3] [7]

03

Understanding the proof has become the bottleneck. The artifact holds about 250,000 declarations, and roughly 60% are unused by the main theorems. It relies on opaque nicknames and has no informal write-up. Reviewers say they do not yet fully understand it, and Martin Hairer is leading an effort to turn it into a readable monograph. [1] [2] [3] [7]

WHY IT MATTERS

The Lean kernel check and axiom audit are public and reproducible, but the existence verdict comes from an OpenAI-appointed panel. If the result holds, AI has produced frontier mathematical physics.

Read the full assessment

Making machine proofs understandable now looks like the costlier step.

Sources are gathered; writing the dossier now.

Executive brief

A committee chaired by Nobel laureate David Gross says a machine-checked proof produced by OpenAI's models establishes that pure four-dimensional Yang–Mills theory exists. That is the "existence" half of a $1 million Clay Millennium Prize problem (Scientific American). The committee also says it does not yet fully understand the proof. The Lean code has about 250,000 declarations, and roughly 60% of them are not needed for the main theorems (GitHub repo). The harder "mass gap" half is still unsolved. OpenAI chose the reviewers itself, and no review independent of OpenAI exists yet.

What changed and event timeline

  1. Natural-language proof produced

    OpenAI says its models wrote an informal Yang–Mills existence proof on this date. A version covering QCD (adding quarks) followed on Sep 18 ().

  2. Lean formalization finished

    OpenAI reports that the formal Lean 4 proof of Yang–Mills existence was finished on this date ().

  3. Public claim

    OpenAI announced the Yang–Mills result together with a claimed disproof of the Exponential Time Hypothesis and a Kakeya-related result. The physics materials stayed with a five-person shepherding group, and OpenAI said it would not claim the prize ().

  4. Outside skepticism

    The 4gravitons blog pointed out that outside reviewers had almost nothing to examine. It also questioned how the shepherds were chosen and asked who has priority, given parallel work by Anthropic-affiliated researchers ().

  5. Committee affirms existence and releases the code

    The Gross committee said the result proves pure Yang–Mills existence under the standard problem statements, then published the Lean code and a validation document (;).

Capabilities and access

  • OpenAI credits "its models" with the result. The committee says it used OpenAI's Astra to help digest the Lean code (Scientific American). Some outlets name the model "GPT-6 Astra" (Pasquale Pillitteri).
  • The Lean 4.32.0 artifact is public on GitHub (repo). The 1,147-page informal manuscript has not been released (Pasquale Pillitteri).
Read the full section
  • OpenAI credits "its models" with the result. The committee says it used OpenAI's Astra to help digest the Lean code (Scientific American). Some outlets name the model "GPT-6 Astra" (Pasquale Pillitteri). The official attribution of which model produced the proof is not stated in the reviewed sources.
  • The Lean 4.32.0 artifact is public on GitHub (repo). The 1,147-page informal manuscript has not been released (Pasquale Pillitteri).

Technical analysis for researchers and developers

  • Approach: The proof follows Bałaban's multiscale renormalization-group program for lattice gauge theory.
  • Main theorem: YangMills.jointExistence builds continuum limits that satisfy the Wightman axioms W0–W3, with positive energy and a unique vacuum.
  • Audit: 17 headline theorems rely only on the three standard axioms (propositional extensionality, classical choice, quotient soundness). The audit found no sorry placeholders.
Read the full section
  • Approach: The proof follows Bałaban's multiscale renormalization-group program for lattice gauge theory. The ultraviolet construction (block-spin steps, gauge fixing, cluster expansions, coupling flow) makes up about 70% of the code (Armstrong informalization). The video transcript describes this as tracking several lattice scales at once so that fine-scale fluctuations stay controlled (video, 2:50).
  • Main theorem: YangMills.jointExistence builds continuum limits that satisfy the Wightman axioms W0–W3, with positive energy and a unique vacuum. It also shows the theory is not free (not a sum of free fields), using stress-tensor correlators (repo).
  • Audit: 17 headline theorems rely only on the three standard axioms (propositional extensionality, classical choice, quotient soundness). The audit found no sorry placeholders. Building the 9,470 modules takes about 184 minutes (Pasquale Pillitteri).
  • Readability: The code is about 1 million lines in 9,477 files. It uses opaque nicknames such as "Treasury" and "Jewel" and comes with no informal explanation (Armstrong).

Claims and evidence

  • Existence of pure 4D Yang–Mills: affirmed by the Gross committee (Scientific American). OpenAI appointed this committee, so the affirmation is not independent of the company.
  • Formal correctness: the Lean kernel check, the axiom audit and the declaration counts (250,319 total, 100,168 used) are published and can be reproduced (repo).
  • QCD extension: reported by OpenAI only. No public artifact backs it (TechTimes).
Read the full section
  • Existence of pure 4D Yang–Mills: affirmed by the Gross committee (Scientific American). OpenAI appointed this committee, so the affirmation is not independent of the company.
  • Formal correctness: the Lean kernel check, the axiom audit and the declaration counts (250,319 total, 100,168 used) are published and can be reproduced (repo). The video's "60% unused" figure (4:17) agrees with these counts.
  • QCD extension: reported by OpenAI only. No public artifact backs it (TechTimes).
  • Corroboration: no review by a group without ties to OpenAI is available.

Context and prior work

  • The Clay problem has two parts: show that quantum Yang–Mills theory exists, and show it has a mass gap (Wikipedia).
  • Bałaban's lattice program from the 1980s laid the groundwork that this proof extends (Pasquale Pillitteri).
  • Gross won his 2004 Nobel for asymptotic freedom, the property the construction depends on. TechTimes presents this as a conflict of interest (TechTimes).
Read the full section
  • The Clay problem has two parts: show that quantum Yang–Mills theory exists, and show it has a mass gap (Wikipedia).
  • Bałaban's lattice program from the 1980s laid the groundwork that this proof extends (Pasquale Pillitteri).
  • Gross won his 2004 Nobel for asymptotic freedom, the property the construction depends on. TechTimes presents this as a conflict of interest (TechTimes).
  • Martin Hairer is leading a crowd-sourced effort to turn the Lean code into a human-readable monograph (Armstrong).

Limitations, safety and contested findings

  • Gaps in what is proved: the proof establishes only subsequential infinite-volume limits, not a unique limit, and does not establish a mass gap (Pasquale Pillitteri; repo).
  • Reviewers' own limits: the reviewers say they do not fully understand the proof, and they used OpenAI's own model to read it (Scientific American).
  • Process criticism: NYU mathematician Scott Armstrong called the shepherding arrangement "needless gatekeeping" (Scientific American).
Read the full section
  • Gaps in what is proved: the proof establishes only subsequential infinite-volume limits, not a unique limit, and does not establish a mass gap (Pasquale Pillitteri; repo).
  • Reviewers' own limits: the reviewers say they do not fully understand the proof, and they used OpenAI's own model to read it (Scientific American).
  • Process criticism: NYU mathematician Scott Armstrong called the shepherding arrangement "needless gatekeeping" (Scientific American).
  • Formal vs. intended statement: Lean can confirm that the logic holds, but not that the formal statement matches the intended problem (TechTimes). The video transcript makes the same point (3:40) and calls the code "spaghetti" dumped on scientists (5:12).

Business and practitioner implications

  • Verification is now the bottleneck, not generation. Machine-checked output at this scale still needs months of expert work to make it understandable.
  • For teams using AI for formal methods: dead code, opaque names and missing informal write-ups add heavily to review cost.
  • Vendor-run review is not independent review.
Read the full section
  • Verification is now the bottleneck, not generation. Machine-checked output at this scale still needs months of expert work to make it understandable.
  • Hygiene matters. For teams using AI for formal methods: dead code, opaque names and missing informal write-ups add heavily to review cost. Requiring pruning and documentation before release is a cheap fix.
  • Vendor-run review is not independent review. Claims vetted by a panel the vendor appointed, using the vendor's own tools, should be read as provisional until outside groups reproduce them.
FOLLOW THE EVIDENCE

The source trail.

Sources (8)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.