GPT-6 Astra Led the Proof: Physicist Nathaniel Craig’s Candidate Lattice Definition of the Standard Model

In the first guest post on OpenAI's new science blog, UC Santa Barbara physicist Nathaniel Craig says GPT-6 Astra turned his idea into a construction and led the detailed proof for a candidate lattice definition of the Standard Model, backed by a Lean formalization. He calls it a beginning, not an end.

Christopher Ort
By Christopher Ort · 7 min read
Share
A university office at night with a chalkboard full of blurred equations and a grid sketch above a desk stacked with papers.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. Background: why the Standard Model has no full definition
  5. How the construction works
  6. Pros and cons
  7. What X is saying
  8. Critical analysis
  9. What's still open
  10. About the analysis
  11. i10x Perspective

OpenAI has opened a science blog, and its first guest post aims high. UC Santa Barbara physicist Nathaniel Craig, a visiting scientist at OpenAI, writes that he worked with GPT-6 Astra on a candidate lattice definition of the Standard Model, the theory that describes every known elementary particle. Craig says the model “took the lead in developing the detailed proof” for a problem that has been open for more than fifty years. He is also careful to call the result a candidate, not a solved case.

i10X Quick Take

Quick Take

  1. What happened: On November 15, OpenAI published Craig’s guest post explaining his October preprint, A gauge-invariant measure for lattice chiral fermions, prepared with the aid of GPT-6 Astra.
  2. The claim: A candidate nonperturbative lattice definition of the Standard Model, built with overlap fermions and a new way of fixing a troublesome phase.
  3. The AI’s role: Craig proposed the refinement strategy; Astra helped turn it into a construction, showed convergence and “took the lead” on the detailed proof. A Lean formalization accompanies the paper.
  4. Human checks: Lattice field theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir reviewed Astra’s proof and asked probing questions, Craig writes.
  5. Still open: Reflection positivity at finite lattice spacing, practical use in simulations and broader expert scrutiny.

In plain English

Physicists have a hugely successful theory of the smallest building blocks of nature, but they have never been able to write it down in a complete, exact form. One way to try is to imagine space as a fine grid and build the theory on that grid. For one of nature’s forces this has stubbornly failed for decades. A physicist now says OpenAI’s newest AI model did most of the hard proof work for a possible fix. Other experts still need to check whether it really works.

What happened

Craig leads the particle theory group at UC Santa Barbara and is a permanent member of its Kavli Institute for Theoretical Physics. In the OpenAI post he writes: “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. Astra helped turn that strategy into a construction and took the lead in developing the detailed proof.”

The underlying paper went to arXiv in October. Its abstract says it constructs “a gauge-invariant measure for overlap fermions on four-dimensional Euclidean lattices in a class of nonabelian chiral gauge theories that includes the Standard Model,” and calls the result “a candidate nonperturbative lattice definition of the Standard Model.”

OpenAI announced the post on X as “the first post on our new science blog.” It comes one week after OpenAI’s Navier-Stokes proof claims and a string of AI-for-math results the company has been promoting.

Background: why the Standard Model has no full definition

Most Standard Model predictions use perturbation theory: switch interactions off, solve exactly, then add interactions back as small corrections. That works brilliantly for many things. It fails for others, like the proton’s mass, which comes mostly from the strong force itself, as Craig explains.

The standard workaround is lattice field theory: replace continuous spacetime with a grid. That approach made the strong force computable from first principles. But the full Standard Model is “chiral”: the weak force treats left-handed and right-handed particles differently. On a naive lattice, trying to put a left-handed fermion on the grid gives “a right-handed partner you never asked for,” Craig writes. Physicists call these unwanted extras doublers.

Decades of work chipped away at the problem. Craig credits Paul Ginsparg and Ken Wilson for the lattice rules, Rajamani Narayanan and Herbert Neuberger for overlap fermions, Martin Lüscher for separating left and right-handed overlap fermions and solving a simpler chiral theory, and Yoshio Kikukawa and collaborators for work toward the weak interactions. One piece stayed missing: how to fix a phase in the fermion measure for the full Standard Model, a question that “remained open for more than 25 years,” according to the post.

How the construction works

Craig’s idea, in his own simplified words: take the force fields on the grid you care about, represent the same fields on a series of finer grids, and compare the allowed fermion states between neighboring grids. Each comparison gives part of the phase correction. Add them up and you get the phase rule on the original grid.

According to the post and the paper:

  • Astra’s contribution: It “suggested the right way to carry a given set of force fields onto finer grids, showed that the comparisons on finer grids would converge to a definite answer, and took the lead in developing the detailed proof.”
  • Scope: The paper treats an SU(5) construction and maps the Standard Model’s gauge fields into it, covering one generation of fermions and, the paper says, three generations by direct generalization.
  • Conditions: The result holds for “admissible” gauge fields, meaning a bound on field strengths at the scale of the grid.
  • Formal checks: The Zenodo record includes a mathematical companion, a Lean verification report and the full Lean source project.

The paper’s acknowledgment, as quoted by Columbia mathematician Peter Woit on his Not Even Wrong blog, is unusually candid: “The manuscript combines author-written sections with extensive rewrites of the Astra-composed proof.”

Pros and cons

For users

  • Pro: A clear public example of an AI model contributing real, checkable mathematics to a famous open problem, not just summarizing papers.
  • Con: Headlines like “AI solves 50-year physics problem” overstate a result its own author calls “a candidate” and “far more an answer in principle than an answer in practice.”

For developers

  • Pro: The workflow is replicable: human strategy, model-driven construction, formal verification in Lean, then expert review.
  • Con: Lean checks only what was formalized. Whether the formal statements fully match the physics claims is exactly the kind of gap critics raised over OpenAI’s Navier-Stokes proof.

For companies

  • Pro: OpenAI gains a showcase for GPT-6 Astra in science, plus a new channel, its science blog, for researchers to publish AI-assisted work.
  • Con: A company blog is not peer review. The credibility of the program depends on independent physicists confirming results.

What X is saying

OpenAI’s announcement drew over 2,400 likes and hundreds of replies:

Enthusiasts called it a breakthrough:

Others focused on verification. “I’d love to read how he checked the model’s steps, since verification is the hard part in physics and software alike,” replied @HKTECH_AI. “If the details hold up under peer review, that’s a real milestone for AI in physics,” wrote @jaydenliSH. And @JamesZhangpro listed the caveats: “Expert scrutiny, reflection positivity and practical simulations are still unfinished.”

Critical analysis

Two things can be true at once. The result is impressive: a long-standing technical gap in lattice field theory, a concrete construction, formal verification artifacts and named outside experts who engaged with it. And the result is early: it is a preprint, its author says scientific dialogue “is required to decide if this prescription has all the properties we desire,” and he is still working on reflection positivity, a condition that connects the lattice model to a quantum theory where probabilities are conserved.

The division of labor matters too. Craig had the core idea. Astra did the heavy lifting of turning it into a proof, by Craig’s account, and the manuscript includes “extensive rewrites of the Astra-composed proof.” That is a different claim from “AI discovered new physics.” It is closer to an extremely capable collaborator executing a hard mathematical program under expert direction.

Note also the source. The story comes from OpenAI’s own blog, written by a visiting OpenAI scientist, about OpenAI’s model. That does not make it wrong, and the arXiv paper and Lean files are public. But independent confirmation from lattice theorists outside the project is what will turn “candidate” into “accepted.”

What’s still open

  • Whether the construction satisfies reflection positivity at finite lattice spacing, which Craig names as his current focus.
  • How to make the method usable in realistic lattice simulations.
  • Defining the interacting theory in continuous spacetime, which the post links to OpenAI’s separate Yang-Mills work.
  • Formal peer review and independent checks of the Lean formalization against the paper’s claims.
  • How OpenAI’s science blog will handle disclosure and review for future AI-assisted results.

About the analysis

Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.

This independent i10x analysis is based on Nathaniel Craig’s guest post on OpenAI’s science blog, his arXiv preprint and the Zenodo record with the Lean formalization, commentary on Peter Woit’s Not Even Wrong blog, and public posts on X. We did not independently verify the mathematics.

i10x Perspective

This is the most interesting kind of AI science story: not “the model did everything,” but “a top physicist had an idea he could not finish, and the model finished the proof under his supervision.” If it holds up, it is a template for how frontier models will change research, by compressing years of technical grinding into weeks. The honest framing from Craig himself is the best part. The worst outcome would be for marketing to round “candidate” up to “solved” before the lattice community has had its say.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Timeline

What happened before

  1. Release

    GPT-6.1 Terra

    1.05M context · $1 / $5 per 1M tokens

    openai.com (source, opens in a new tab)
  2. Acquisition

    OpenAI acquires Orchestra Bio

    Acquires Orchestra Bio

    openai.com (source, opens in a new tab)
  3. Release

    GPT-6.1 Sol

    1.05M context · $2 / $10 per 1M tokens

    developers.openai.com (source, opens in a new tab)
  4. Release

    GPT-6 Sol & Luna

    1.05M context · $2 / $10 per 1M tokens

    Two GPT-6 reasoning models. Price shown is GPT-6 Sol (≤272K prompt); GPT-6 Luna costs $0.10 / $0.50.

    developers.openai.com (source, opens in a new tab)
  5. Release

    GPT-6 Astra

    1.05M context · $10 / $50 per 1M tokens

    OpenAI's most capable model for reasoning, coding, computer use and research.

    openai.com (source, opens in a new tab)
Full OpenAI timeline ›

Editor's picks

  1. CVE-2026-129877: DeepAudit exposes LLM API keys and Git tokens to unauthenticated attackers
  2. CVE-2026-129875: REBUILD 4.5.0–4.5.4 password hash disclosure via MCP QueryRecords
  3. South Korea Weighs Nuclear Power to Meet AI Electricity Demand
  4. UK and US Partner to Accelerate Nuclear Fusion with AI
  5. Texas and New Mexico Data Center Emissions Rise With AI Buildout