SuperIntelligence Infrastructure 7.2%reading
researchOpenAILayer lab

OpenAI Promotes a GPT-6 Astra Lattice Proof That Edward Witten Reworked With Claude 12 Days Earlier

OpenAI's Nov. 15 post credits GPT-6 Astra with leading Nathaniel Craig's lattice proof; Edward Witten and Stephen Hsu had already built on it, and Peter Woit calls the advance modest.

▲ +0.3 Capability reported Reading after Nov. 15, 2026 (2 pieces that day): 7.5

By Ryan Elliott Dennis · 10 sources · 9 min read

Glass office building at 1515 Third Street in Mission Bay, San Francisco, home to OpenAI
OpenAI's headquarters at 1515 Third Street in San Francisco, photographed June 15, 2025; the company opened its science blog on Nov. 15 with Nathaniel Craig's guest post. Photo: Coolcaesar, Wikimedia Commons, CC BY 4.0

OpenAI opened a science blog on Nov. 15, 2026 with a guest post by Nathaniel Craig, who leads the particle theory group at the University of California, Santa Barbara and is a visiting scientist at OpenAI 79. Craig's post credits GPT-6 Astra, OpenAI's model, with leading the detailed proof in his Oct. 6 preprint, a candidate way to define the Standard Model of particle physics on a lattice 15. Forty days separate the preprint from the post. Edward Witten of the Institute for Advanced Study posted a 33-page variant of the construction on Nov. 3, day 28, and thanked Anthropic's Claude Opus 5.5 for help 2.

Why should a reader outside physics care? OpenAI's claim that its model did research counts for little on this ledger until someone outside OpenAI works with the result. On Oct. 26 the ledger credited OpenAI where Lean, a program that checks proofs line by line, and a rival lab's matching result stood behind a claim. Here the outside evidence is a pair of papers by physicists outside the project, and one of them leaned on a competitor's model.

Outside use of GPT-6 Astra's lattice proof adds 0.3 to capability

The move
Capability steps up 0.3 at reported confidence, the floor of the 0.3 to 0.8 band: Edward Witten and Stephen Hsu built on a proof GPT-6 Astra led, while a referee's report and two physics problems stay open.
The data5 rows · sources
MeasureValue
Reading before this day7.2
This piece's move+0.3 (Capability, reported)
Band for reported evidence0.3 to 0.8
Reading after the day (with 1 other piece that day)7.5
Distance to 10092.5

What did Craig claim, and who wrote the proof?

Physicists compute with the Standard Model by approximation, and a lattice, a grid of points standing in for spacetime, offers a route to an exact definition 8. One obstacle has blocked that route for decades: the Standard Model is chiral, which means its left-handed and right-handed particles carry different charges, and a grid tends to add unwanted copies of each particle. Physicists call this the fermion doubling problem 2. Martin Lüscher wrote a prescription for simpler theories, and the full Standard Model case stayed open for more than 25 years 5.

Craig's preprint, "A gauge-invariant measure for lattice chiral fermions," builds the missing piece for a class of theories that includes the Standard Model 1. Witten's summary of the method is the shortest one in print: "Craig compares chiral fermions on ever finer lattices, ultimately defining the fermion measure via a continuum limit." 2 The measure is the rule for summing over every state the particles can take.

Who did the work? Craig's post divides it. "I had a strategy for tackling this problem, but had not been able to prove that it worked in all the cases it needed to cover," he writes, in a passage AI Weekly reproduced on Nov. 16 5. His next sentence hands over the rest: "Astra helped turn that strategy into a construction and took the lead in developing the detailed proof." The paper's acknowledgments go further: "OpenAI's GPT-6 Astra model was essential to this result." 1

Read the nouns: Craig keeps "strategy" for himself and gives Astra "construction" and "proof," the two parts a referee checks. That account comes from the paper's one author, a visiting scientist at the company that sells the model.

Kohn Hall, the Kavli Institute for Theoretical Physics building on the UC Santa Barbara campus
Kohn Hall, home of the Kavli Institute for Theoretical Physics at UC Santa Barbara, in October 2006; Craig's preprint lists the institute as his affiliation. Photo: Jonathan Pfeiffer from United States of America, Wikimedia Commons, CC BY 2.0

Two physicists, two labs' models, one lattice construction

Who connects
Craig credits OpenAI's GPT-6 Astra with leading his proof, Witten thanks Anthropic's Claude Opus 5.5 for help on his variant, and Hsu and Woit each answered Craig's paper within 15 days.Sources [2] [3] [4] [5] [7]
The data6 rows · sources
FromLinkToSource
GPT-6 Astraled the detailed proofNathaniel Craig[5]
OpenAIhosts as visiting scientistNathaniel Craig[7]
Edward Wittenpresents a variantNathaniel Craig[2]
Claude Opus 5.5assisted on many detailsEdward Witten[2]
Stephen D.H. Hsustarts from his constructionNathaniel Craig[3]
Peter Woitcalls the advance modestNathaniel Craig[4]

Which outside physicists have worked with the construction?

Two have, and both published before OpenAI's post. Stephen D.H. Hsu of Michigan State University posted a paper on Oct. 21, 15 days after Craig, and revised it on Nov. 6 to 113 pages 3. Its introduction opens from "Craig's recent construction of smooth, gauge-covariant connections from admissible lattice links". Hsu keeps Craig's method for rebuilding the force fields from the grid and treats the particles a different way.

Witten followed on Nov. 3 with "A Note On Lattice Definition of Gauge Theories With Chiral Fermions" 2. His verdict on Craig's paper sits in the introduction: "Very recently, Craig overcame the difficulties and, with the aid of several new ideas, finally constructed a lattice regularization of chiral gauge theories in enough generality to encompass the Standard Model." The adverb "finally" belongs to a physicist who watched the problem stay open, and Witten then presents a variant of his own, one Hsu had proposed as well.

Machine help appears in Witten's note too. "Assistance on many details was provided by Claude Opus 5.5," his acknowledgments read, and the text credits Anthropic's model with suggesting a smooth way to fill in the fields between grid points. OpenAI's model led a proof in October, and Witten, generalizing the idea in November, did so with Anthropic's. Both papers stop short of saying their authors checked Astra's proof line by line; each builds its own route to a similar result.

Trade coverage of OpenAI's post missed both papers. AI Weekly's Alexis Dufresne wrote on Nov. 16 that "no independent published assessment has yet appeared" 5. MLQ.ai went further: "MLQ.ai found no published reaction from physicists outside the project as of Sunday evening U.S. time." 6 Witten's note had been on arXiv for 12 days by that Sunday.

Fuld Hall, the red-brick main building of the Institute for Advanced Study in Princeton
Fuld Hall at the Institute for Advanced Study in Princeton, New Jersey, in July 2023; Edward Witten's Nov. 3 note carries the institute's Einstein Drive address. Photo: Zeete, Wikimedia Commons, CC BY-SA 4.0

Witten's variant reached arXiv 12 days before OpenAI's post

Timeline
Craig posted on Oct. 6, Woit answered on Oct. 8, Hsu built on the construction on Oct. 21 and Witten on Nov. 3; OpenAI's post of Nov. 15 arrived 40 days after the preprint.Sources [1] [2] [3] [4] [7] [9]
The data6 rows · sources
DateEventSource
Craig submits arXiv:2610.09093, crediting GPT-6 Astra[1]
Woit calls the result a modest advance over Lüscher[4]
Hsu posts a paper starting from Craig's construction[3]
Witten posts a 33-page variant, thanking Claude Opus 5.5[2]
Hsu revises his paper to 113 pages[3]
OpenAI opens its science blog with Craig's guest post[7] [9]

Why does Peter Woit call the advance modest?

Peter Woit, a Columbia University mathematician who writes the blog "Not Even Wrong," wrote about the preprint on Oct. 8, two days after Craig submitted it 49. "The results of this work are a modest advance over Lüscher, still a long ways from providing a full definition in principle of a theory with the desired features in a continuum limit," he wrote. A second sentence limits his own reading: "I haven't tried to follow the details of the paper, especially not the complicated ones provided by Astra."

Woit objects to the usefulness of the result more than to its correctness. In a comment under the post he wrote: "My main worry about this kind of construction is that, whether or not it solves problems of principle, it doesn't seem to provide a way forward to actually do calculations." 4

Witten's note agrees with Woit on that point. "In any dimension and for any anomaly-free representation, this approach can be implemented on a computer in principle, but not in practice by Monte Carlo methods because the chiral fermion path integral is not positive," his abstract says 2. Monte Carlo sampling is how lattice physicists get numbers out of a supercomputer, so the definition arrives ahead of any new prediction. Craig concedes as much, calling the method "far more an answer in principle than an answer in practice" 5.

One line of Woit's post aged within a month. "Bringing AI firepower to bear on pushing forward a complicated proposal that was stalled for good reason isn't likely to lead anywhere interesting," he wrote 4. Twenty-six days later Witten called the difficulties overcome and added a paper of his own, though Woit's worry about calculations still stands and Witten shares it.

Witten says Craig overcame the difficulties; Woit calls the advance modest

Both sides
Witten's Nov. 3 note accepts Craig's construction as covering the Standard Model, while Woit's Oct. 8 post calls it a modest advance over Lüscher; the evidence leans slightly toward Witten, who wrote 26 days later.Sources [2] [4]
The data2 rows · sources
SideWhoClaimSource
ForEdward WittenCraig overcame the difficulties and finally constructed a lattice regularization of chiral gauge theories in enough generality to encompass the Standard Model.[2]
AgainstPeter WoitThe results of this work are a modest advance over Lüscher, still a long ways from providing a full definition in principle.[4]

What do the Lean files cover, and what stays open?

Craig's paper describes three checks: a Lean 4 formalization whose 16 modules compiled, a parallel construction in two dimensions, and a second derivation, according to MLQ.ai 6. The acknowledgments then say who ran them: "My colleagues and I proposed tests of the method, including Lean verification, that were carried out by Astra; supplementary materials documenting some of these checks were written by Astra." 1 So the model that led the proof also ran the checks on it and wrote them up. Lean is independent software, and trust enters at the choice of which statements to formalize and at the question of who made that choice.

Craig marks that limit himself in the body of the paper. "This is not to suggest, however, that Lean verification is the final word in physics, where there is a great deal more that can go wrong beyond the failure of a formally stated conclusion to follow from its specified definitions and assumptions," the paper says 1.

Human readers appear in the acknowledgments as well, and some of them share Craig's tie to OpenAI. Craig writes that he is "deeply indebted to Maarten Golterman and Yigal Shamir" and thanks Andreas Kronfeld for comments on the manuscript; all three are lattice field theorists, and their own assessments remain unpublished 15. He also thanks Xiao-Liang Qi, Eva Silverstein and Dan Roberts. Woit reported on Nov. 2 that Roberts is an OpenAI employee and that Qi and Silverstein are OpenAI visiting scientists, as Craig is 10.

Two physics problems stay open by the paper's own list. Its conclusions ask future work to "establish reflection positivity compatible with admissibility" and add: "And, of course, a definition of the interacting continuum theory remains open." 1 Reflection positivity is the condition that lets a lattice theory be read as a quantum theory, and Hsu's abstract reports getting partway there: "We also establish reflection positivity at zero Yukawa coupling and zero theta for a modified gauge-field reconstruction; its continuum equivalence with Craig's original reconstruction is not proved." 3

What outsiders and Lean have checked in Craig's paper, and what waits

The record
Solid tiles rest on two outside papers and 16 compiled Lean modules; dashed tiles hold the step at 0.3: Astra ran its own checks, three readers' views are unpublished, and two physics problems stay open.Sources [1] [2] [3] [5] [6]
The data10 rows · sources
ColumnItemSource
On the recordWitten writes that Craig overcame the difficulties[2]
On the recordHsu builds a 113-page paper on Craig's reconstruction[3]
On the recordSixteen Lean 4 modules compiled[6]
On the recordGolterman, Kronfeld and Shamir read the work for Craig[5]
Claimed or openAstra took the lead on the proof, by Craig's account[5]
Claimed or openAstra carried out the Lean checks on its own proof[1]
Claimed or openAssessments by the three lattice theorists stay unpublished[5]
Claimed or openReflection positivity for Craig's original reconstruction[3]
Claimed or openA definition of the interacting continuum theory[1]
Claimed or openMonte Carlo simulation, which Witten rules out in practice[2]

Why does the step stop at 0.3?

The ledger moves capability up 0.3 at reported confidence, the floor of the 0.3 to 0.8 band. Evidence for a step is plain: within 28 days, Hsu and Witten, both outside the project, built on a proof that GPT-6 Astra led, and Witten wrote that the problem's difficulties were overcome.

Four facts hold the step at the floor. First, the strategy was Craig's, so GPT-6 Astra finished a plan that a human physicist had set and left unproven. Second, the split of labour between Craig and Astra rests on Craig's account alone, written as a visiting scientist at OpenAI. Third, Astra ran the Lean checks on its own proof, and a referee's report is still pending. Fourth, reflection positivity and the continuum theory are open, and Witten says the method resists Monte Carlo simulation.

Compare the ledger's entry for Oct. 26, when a second lab's model reached the same 3-SAT result independently and Lean files were public, which earned 0.4. Witten and Hsu extended Craig's result, a weaker form of confirmation than reproducing it.

What would each side need to be true? Witten's reading holds if a referee, or Golterman, Shamir and Kronfeld in print, confirm the proof. Woit's holds if the construction stays a definition that computes little, or if a gap turns up in the paper's 60 plus 63 pages. For a student deciding whether to trust a model-written proof, the working verdict is this: credit GPT-6 Astra with a proof that outside physicists chose to build on, and treat "took the lead" as Craig's account until an audit of the transcripts shows it.

By the numbers

  • 40 days: the gap between Craig's Oct. 6 preprint and OpenAI's Nov. 15 post 19.
  • Twenty-eight days after the preprint, on Nov. 3, Witten posted a 33-page variant 2.
  • 113 pages: the length of Hsu's paper after its Nov. 6 revision, first posted Oct. 21 3.
  • Sixteen Lean 4 modules compiled for the formalization, per MLQ.ai 6.
  • More than 25 years: the time the Standard Model case stayed open after Lüscher's prescription 5.
  • Craig's preprint runs 60 plus 63 pages with 12 figures 1.
  • Three lattice theorists, Golterman, Kronfeld and Shamir, read the work for Craig, and their assessments are unpublished 5.

What to watch

A referee's report or a journal acceptance for arXiv:2610.09093 would confirm this move, and a published assessment from Golterman, Kronfeld or Shamir would do the same 5. Release of the Astra transcripts, or an outside audit of who wrote which step, would test the phrase "took the lead." A proof of reflection positivity for Craig's original reconstruction would close the gap Hsu named 3. An error found in the main text or in the Lean statements would reverse the step.

Sources

  1. 1A gauge-invariant measure for lattice chiral fermions (arXiv:2610.09093), arXiv, Nathaniel Craig, Oct. 6, 2026
  2. 2A Note On Lattice Definition of Gauge Theories With Chiral Fermions (arXiv:2611.05051), arXiv, Edward Witten, Nov. 3, 2026
  3. 3The Standard Model on a finite bosonic lattice with continuum Weyl fermions (arXiv:2610.29019), arXiv, Stephen D.H. Hsu, Oct. 21, 2026
  4. 4AI and Chiral Fermions, Not Even Wrong, Peter Woit, Oct. 8, 2026
  5. 5OpenAI Says GPT-6 Astra Led Proof for Lattice Standard Model, AI Weekly, Alexis Dufresne, Nov. 16, 2026
  6. 6OpenAI says GPT-6 Astra led the proof behind a candidate lattice definition of the Standard Model, as it launches a science blog, MLQ.ai, MLQ.ai Newsroom, Nov. 16, 2026
  7. 7Physicist credits GPT-6 Astra with leading proof work on Standard Model proposal, Superpower Daily, Superpower Daily Editorial Desk, Nov. 15, 2026
  8. 8GPT-6 Astra Helps Physicist Tackle 25-Year Lattice Problem for the Standard Model, XenoSpectrum, Y Kobayashi, Nov. 16, 2026
  9. 9OpenAI launches a science blog for researchers, CASRAI, CASRAI Editorial Board, Nov. 16, 2026
  10. 10OpenAI Conference at KITP, Not Even Wrong, Peter Woit, Nov. 2, 2026