GPT-6 Astra helped close a 50-year gap in the Standard Model

OpenAI's science blog posted a candidate lattice definition of the Standard Model, with GPT-6 Astra leading the proof. The 50-year problem has a fix in principle, not yet in practice.

Share
GPT-6 Astra helped close a 50-year gap in the Standard Model

OpenAI's most interesting research post this month is not about a model. It is about a hole in particle physics that has been open since the Standard Model took shape — and the language model that helped a physicist write the proof.

On November 15, OpenAI published the first guest post on its new science blog: a candidate nonperturbative lattice definition of the Standard Model, built by UC Santa Barbara physicist Nathaniel Craig with GPT-6 Astra. The paper behind it went up on arXiv six weeks earlier, and it is unusually checkable for a claim this large — a 60-page paper, a 63-page mathematical companion, three supplementary documents, and a Lean formalization of the mathematics, posted by Craig. Craig is not a visiting blogger with a nice result. He runs UCSB's particle theory group and holds a permanent membership at its Kavli Institute for Theoretical Physics. GPT-6 Astra is a collaborator on the proof, not an author on the paper.

What was actually missing

The Standard Model describes the known elementary particles and three of the four forces, and its predictions match experiment across eighteen orders of magnitude. Nearly all of that precision comes from perturbation theory: pretend the interactions are switched off, solve that exactly, then add the interactions back as progressively smaller corrections. It works beautifully — it predicts the electron's magnetism to within a few parts in a trillion.

It falls apart exactly where the interesting physics lives. A proton's mass comes mostly from the quarks and gluons churning inside it; switch off the strong force and the proton simply falls apart, so there is nothing small left to expand in. To handle that, physicists put the theory on a lattice: discretize spacetime into a grid, which makes the theory finite, rigorous at that grid spacing, and computable on a machine. Lattice field theory is why anyone can derive the proton's mass from first principles.

The obstruction to doing that for the whole Standard Model is chirality. The weak force treats left- and right-handed fermions differently, so the theory does not look the same in a mirror. Put a left-handed fermion on a lattice the naive way and the rules hand you a right-handed partner you never asked for — a "doubler" — which gives the weak force the wrong particle content. Overlap fermions, built on work by Paul Ginsparg and Ken Wilson and realized by Rajamani Narayanan and Herbert Neuberger, get rid of the doublers. Martin Lüscher then showed, in 1999, how to assign the remaining phase — the measure that makes the construction gauge-invariant — for simpler chiral theories. For the full Standard Model, that alignment stayed open for more than twenty-five years. In the blog post's own accounting, the full nonperturbative definition has been missing for over fifty.

What Craig and Astra did

The idea is recognizable even to a non-physicist. Take the force fields on your grid, represent the same fields on a succession of finer grids, and compare the allowed fermion states on each pair of neighbours. Adding up the differences recovers the phase correction you need on the original grid — a refinement argument rather than a single closed-form trick.

Craig had that strategy but could not prove it covered every case it needed to. In his account for OpenAI, Astra suggested how to carry a given set of force fields onto finer grids, showed the comparisons on finer grids would converge to a definite answer, and took the lead in developing the detailed proof. The result is a gauge-invariant measure for overlap fermions on four-dimensional Euclidean lattices, covering a class of nonabelian chiral gauge theories that includes the Standard Model, valid on admissible gauge fields under a bound on field strengths.

Why this is a candidate, not a solution

Read the hedges, because they are doing real work. Craig calls the construction "much more a beginning than an end," and "far more an answer in principle than an answer in practice." Two technical gaps are stated plainly: whether the construction satisfies reflection positivity at finite lattice spacing is still open, and defining the interacting theory in continuous spacetime leans on OpenAI's recent proof of the existence of Yang-Mills theory. Neither is a formality — the first is what connects a lattice model to a quantum theory where probabilities behave, and the second is what turns a finite grid into physics.

Outside reviewers did look at it. Three lattice field theorists — Maarten Golterman, Andreas Kronfeld and Yigal Shamir — went through Astra's proof and pushed back on it, which is more scrutiny than most AI-assisted results get before publication. It is not peer review. The field has not voted.

The AI-assisted-physics problem

Claims of this shape deserve a raised eyebrow, and we have covered the reason: Princeton's Mengdi Wang argues LLMs have not yet made a real scientific discovery, because the work that looks like discovery tends to be a human framing a problem and a model filling in steps. This case fits that critique in part — Craig set the strategy — and departs from it in the part that matters, because the missing piece was the proof itself, and the model supplied it.

What separates this from the pile of breathless AI-science announcements is verification. The paper is single-authored and predates the blog by six weeks, so the claim was not manufactured for a launch. There is a Lean formalization and two Zenodo archives of calculation records, so the mathematics can be machine-checked rather than argued about. Named experts outside OpenAI and outside the author's institution reviewed it and are credited.

The test now is boring and decisive: whether independent lattice groups pick the construction up, whether it survives a journal, and whether reflection positivity holds. If it does, a fifty-year-old definitional hole has a fix, and the interesting question becomes how many more of them a language model can be pointed at. If it does not, OpenAI will have published a physics preprint on its own blog that the field quietly declined to use — and that, too, would be a fact worth knowing about how AI-assisted research gets announced.

There is a second story here, and it is about the channel. OpenAI now hosts a science blog where a working academic publishes a guest result, launched the same day as the post. That follows its run of mathematics results — the last FrontierMath tier, a formalized proof, a disproved conjecture — and it puts research credibility, not product, in front of the audience. Compare it with how the other labs publish: DeepMind's results arrive through journals and preprint servers with a corporate blog attached afterwards. OpenAI is trying to own the venue as well as the result.

What to watch

Watch the arXiv version history and the citation count over the next two months — that is where you will see whether lattice people engage or shrug. Watch whether the reflection-positivity question gets answered in a follow-up, since that is the specific technical objection most likely to sink the construction. And watch the science blog: if the next guest post is another result of this weight, OpenAI has built something that functions like a journal with a brand on it.

A fifty-year problem, a model that wrote the proof, and reviewers who were not obliged to be polite. If that becomes routine, what should count as an AI discovery — the result, or the fact that a machine got there first? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Gemini 4 Hydrogen: $1.50 in, $7.50 out per million tokens

Gemini 4 Hydrogen: $1.50 in, $7.50 out per million tokens

Google put a Gemini 4 model on every free surface it owns today, and priced it below everything else in the family. The benchmark is fine. The price is the story. Google launched Gemini 4 Hydrogen today at $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1 million-token context window with no long-context surcharge — the cheapest tier in the Gemini 4 line, and 25% below the introductory rate Google is still charging for Gemini 4 Argon, according to the launch…

Nvidia GR00T 2: robots learn tasks from human video, no teleop

Nvidia GR00T 2: robots learn tasks from human video, no teleop

Nvidia's robotics lab stopped mapping camera pixels straight to motor commands and started predicting the next second of video instead. That single change is why a robot can now do a job it has never been taught. Nvidia Research published GR00T 2, a robot foundation model that learns manipulation tasks from video of people and then performs them on robot bodies that never saw a demonstration of the task — with no teleoperation at all.

Long context costs 3.6x to 7x more per token. Compaction holds up

Long context costs 3.6x to 7x more per token. Compaction holds up

Two numbers a buyer never sees on a pricing page: what a long prompt actually costs the seller, and how little accuracy the buyer gives up by not sending one. Serving an open-weight model at a 256K-token context instead of 16K costs 3.6 times more per generated token for GLM 5.2, 5.6 times more for Kimi K2.6 and 7 times more for Qwen 3 Coder Next, according to a 31-page NVIDIA technical report posted to arXiv on November 3.

Arete-27B: Nous cut Qwen's propaganda rate from 22% to 0.9%

Arete-27B: Nous cut Qwen's propaganda rate from 22% to 0.9%

Nous Research took Alibaba's Qwen3.8-27B, put about fifty million tokens of supervision through it, and shipped back a model that answers political questions like a Western lab's build. The interesting part is how little it took — and that the whole thing is checkable. Nous Research has released Arete-27B, an Apache 2.0 model post-trained from Alibaba's Qwen3.8-27B, and says the share of answers containing Chinese state propaganda fell from 22 percent to 0.9 percent while general capability…

Meta Muse gets Active Memory: 90% on Meta's memory eval, up from 76%

Meta Muse gets Active Memory: 90% on Meta's memory eval, up from 76%

Meta stopped selling its frontier model to most people and started using it as a teacher instead. On Friday the results of that swap shipped to everyone with the app. Meta began rolling out new models to every person using Muse and Meta AI on Friday — not Muse Wonder, its frontier model, but versions of its existing consumer models trained using Wonder, the company said in its announcement. Paying subscribers still get Wonder itself for harder tasks.

Inkling 2 Small: 594B parameters, 31 on the AI index

Inkling 2 Small: 594B parameters, 31 on the AI index

Two things happened in the same twelve hours: a US lab shipped its best model yet and promised to give the weights away, and the market decided the AI trade had gone too far. Both are worth your morning. Thinking Machines Lab released Inkling 2 Small on Thursday — a 594-billion-parameter mixture-of-experts model that activates 17 billion parameters per token and scores 31 on the Artificial Analysis Intelligence Index, the highest mark any US lab has posted for a model it says it will open up.