On November 15, 2026, OpenAI published the first guest post on its new science blog. In it, UC Santa Barbara physicist Nathaniel Craig, a visiting scientist at OpenAI, says he used GPT-6 Astra to turn a strategy into a proof. The target is a candidate nonperturbative lattice definition of the Standard Model of particle physics. OpenAI's post says Astra "took the lead in developing the detailed proof."
Craig is careful. He calls the result "a candidate" and "far more an answer in principle than an answer in practice." This post explains the physics in plain language, what the AI is said to have done, and what stays open.
TL;DR
| Question | Answer |
|---|---|
| What is claimed? | A candidate lattice definition of the full Standard Model, with a detailed proof led by GPT-6 Astra |
| Who says so? | Nathaniel Craig (UCSB, visiting scientist at OpenAI), on OpenAI's science blog |
| What problem is this? | Giving the Standard Model a nonperturbative definition, open for more than 50 years |
| Is it settled? | No. Craig says expert dialogue must decide if the prescription has all the properties physicists need |
| Who checked it? | Maarten Golterman, Andreas Kronfeld and Yigal Shamir looked over the proof, per Craig |
| Source | OpenAI science blog, plus an October preprint Craig links |
What does it mean to "define" the Standard Model?
The Standard Model describes the known elementary particles and three of the four fundamental forces: the strong force, the weak force and electromagnetism. Craig notes that its predictions agree with experiment across eighteen orders of magnitude.
Yet a foundational gap remains. Physicists usually compute with perturbation theory. They imagine the interactions turned off, solve that case exactly, and add the interactions back as a series of small corrections. That method predicts the strength of the electron's magnetism to within a few parts in a trillion.
It fails when the interactions are not small. Craig's example is the proton. Its mass comes mainly from the motion and interactions of quarks and gluons. Turn off the strong force and the proton falls apart, so perturbation theory is a poor guide to the proton's mass.
For the full Standard Model in continuous spacetime, Craig writes that we lack a nonperturbative definition, "or even an understanding of whether the interacting theory is mathematically well defined."
Why a lattice?
A lattice replaces smooth spacetime with a grid of points. The grid spacing sets the shortest distance the theory can resolve. A finite grid gives a rigorous definition at that spacing, and it lets physicists put the theory on a computer and calculate directly.
This approach, lattice field theory, has worked well for the strong force. Craig credits it with a first-principles understanding of the proton mass.
Pendulum on a physics bench, a visual stand-in for the Standard Model on a lattice problem
The obstacle: the weak force is chiral
The weak force treats left-handed and right-handed particles differently. A theory with that property is chiral. Craig uses gloves in a mirror: a left glove becomes a right glove, so a world with only left gloves looks different when reflected.
The strong force and electromagnetism treat both hands the same, so they fit on a lattice more easily. The weak force does not. Craig explains that if you put a left-handed fermion on the lattice in the most natural way, you get a right-handed partner you never asked for. This is fermion doubling. A naive lattice Standard Model therefore has the wrong particle content for the weak force, and that is one reason a full definition has been missing.
Craig's mirror analogy
To explain doubling, Craig uses light in a box. Send left-circular light into a box with ordinary mirrors. Each bounce flips it to right-circular, so the box ends up with both polarizations. The box supplied the other hand.
The fix is a chiral mirror, a film whose molecules twist in a spiral. It sends left-handed light back left-handed. Craig says some scarab beetle shells do the same thing.
The lattice has a matching fix, built over decades. In his account:
- Paul Ginsparg and Ken Wilson (1982) worked out the lattice rules analogous to what chiral mirrors need.
- Rajamani Narayanan and Herbert Neuberger (1994) built the analog of the chiral mirror: overlap fermions.
- Martin Lüscher (1999) showed how to tell left-handed from right-handed overlap fermions.
- Yoshio Kikukawa and collaborators explored how this could describe the weak interactions.
For background on this line of work, see the recent review of the Ginsparg-Wilson relation and overlap fermions, which calls a lattice-regulated chiral gauge theory a long-standing open problem.
The missing piece: getting the phase right
Overlap fermions keep only the left-handed states you want. But one piece was missing: the phase.
Craig's water-wave picture helps. Two waves arriving at a dock piling in phase lift the water higher. A crest meeting a trough cancels out. In the mirror box, the returning light must arrive in step with the laser, so you rotate one mirror to fix the phase.
On the lattice, the phase depends on what the force fields do at every grid point. Take the force fields on a loop of small changes that returns to the start. The same fermion states are allowed at the end, yet the phase can differ. You must compensate for the phase picked up around every loop. Craig writes that Lüscher worked this out for simpler theories, such as a chiral version of electromagnetism, but that the case for the full Standard Model "remained open for more than 25 years."
What Craig proposed, and what Astra did
Craig's idea was to represent the same force fields on a succession of finer grids. Comparing the allowed fermion states on two successive grids gives one part of the phase correction for a small change. Adding the parts up gives a phase rule on the original grid. The proof must show this rule gives the right correction around every allowed loop.
The construction supplies the fermion measure for the Standard Model at finite lattice spacing, under a bound on field strengths at the scale of the grid. That caveat matters: it is not the unrestricted theory.
On Astra's role, the post is specific. GPT-6 Astra:
- Suggested the right way to carry a given set of force fields onto finer grids.
- Showed that the comparisons on finer grids converge to a definite answer.
- Took the lead in developing the detailed proof.
Craig adds that this took "several rounds of work with Astra and discussions with outside experts." He had a strategy but "had not been able to prove that it worked in all the cases it needed to cover."
Lab flask illustration for AI-assisted scientific research with GPT-6 Astra
Who checked it
Craig names three lattice field theorists who reviewed Astra's proof: Maarten Golterman, Andreas Kronfeld and Yigal Shamir. He says they spent time looking it over and asking probing questions that helped build his confidence.
That is not the same as a published verdict. The post does not claim that they endorse the result. It says their questions helped. A reader should treat the claim as credible enough for experts to examine, and not yet accepted.
What comes next, in Craig's words
Craig lists open work:
- Reflection positivity at finite lattice spacing. He says he is thinking about whether the construction satisfies this condition, which helps connect lattice calculations to a quantum theory in which probabilities are conserved.
- Usefulness. Much work lies ahead to make the method practical for realistic lattice simulations.
- The continuum. The construction leaves open the interacting theory in continuous spacetime. He points to OpenAI's recent proof of the existence of Yang-Mills theory as the missing puzzle piece. Per Craig, that let OpenAI prove the existence of the simplest unified Yang-Mills theory that includes the gauge fields and fermions of the Standard Model, starting from overlap fermions on the lattice and taking the continuum limit.
That last claim links to a thread we have covered. See our report on the Gross group's validation of OpenAI's Yang-Mills Lean code, which says the committee confirmed existence of pure Yang-Mills and not the Clay mass gap. Note that the committee statement concerns pure Yang-Mills. The Standard Model extension Craig mentions is a separate claim in OpenAI's post, and we did not find an independent expert statement on it.
OpenAI's new science blog
The lattice post is the first guest post on the OpenAI science blog, which launched the same day. OpenAI says it will publish:
- Research and discoveries from OpenAI, collaborators and the wider community.
- Methods and practical guidance, including "how to check the work and where a method falls short."
- Perspectives from scientists using AI, including open questions about scientific practice.
The launch post recaps recent results OpenAI points to:
- In September, an internal system produced a proof addressing Navier-Stokes existence and smoothness. See our Navier-Stokes agent swarm coverage.
- In early October, a broad set of mathematical results, including one on the quasi-Riemann hypothesis.
- Later in October, results on how waves concentrate, the limits of efficient computation, and the foundations of interlacing quantum field theory. Our October three-result explainer covers that batch.
- In biology, thousands of GPT-Rosalind agents analyzed data from millions of mouse brain cells in 32 minutes with PerturbAI and Scripps Research. For the model's lab-facing side, see our GPT-Rosalind workbench post.
OpenAI also flags the open questions: how new machine-made arguments are checked and explained, how scientists choose among many more hypotheses, and what AI means for training and attribution.
Claimed versus verified
| Item | Status |
|---|---|
| Candidate lattice definition exists as a preprint | Craig says he posted it in October; read the paper through OpenAI's link |
| GPT-6 Astra took the lead on the detailed proof | Craig's account, published by OpenAI; not independently audited |
| Outside experts reviewed it | Craig names three; no public statement from them found by us |
| Result holds under a field-strength bound at the grid scale | Stated by Craig |
| Settles the Standard Model definition | Not claimed. Craig calls it a beginning |
| Continuum limit for the Standard Model | Relies on OpenAI's Yang-Mills result; no independent validation found for the Standard Model case |
We read OpenAI's post and its science-blog launch note. We did not read the preprint line by line, and we cannot judge the mathematics ourselves. Where this post says what the paper proves, it is repeating Craig's description.
Why it matters for people who use AI
First, this is a domain expert's firsthand account, written by the person who steered the work. That is a different kind of evidence from a lab press release. Craig credits the model with real steps (the refinement map, the convergence argument, the proof) and credits humans with the strategy and the checking.
Second, it shows a pattern that repeats across OpenAI's recent science posts: the model generates a long argument, then experts audit it. Our report on Astra's ten math proofs with Lean certificates shows the machine-checked version of that pattern. The lattice proof is a paper argument, so the audit is human. That is slower and less certain.
Third, treat the headline with care. "AI defines the Standard Model" is not what the post says. The post says a candidate definition exists, under a bound, pending expert review. Our Millennium Prize fact-check takes the same approach to grand claims.
What to do if you work in this area
- Read the preprint and the post together. The post is a guide; the paper is the claim.
- Check the bound. The construction holds under a bound on field strengths at the grid scale. Ask whether your use case fits.
- Ask for formal checks where possible. A paper proof from a model needs expert audit. A Lean formalization gives a machine check, as in the Lean cost-collapse discussion.
- Wait for independent commentary. Watch for statements from lattice groups on reflection positivity and on whether the phase rule is canonical.
- Log what the model did. Craig's list of Astra's steps is a good template for crediting AI in a methods section.
Related reading
- OpenAI Yang-Mills Lean validation by the Gross group
- RoundSquare and the mass-gap debate
- OpenAI Astra and ten math proofs with Lean certificates
- OpenAI Navier-Stokes agent swarm
- Did AI solve a Millennium Prize problem? A fact-check
- OpenAI GPT-Rosalind workbench for life sciences
Quotes and details come from OpenAI's November 15, 2026 posts. We update this post when experts publish assessments.
