The Standard Model on the lattice: a preprint proposes a construction, OpenAI claims a role for GPT-6 Astra
On 6 October 2026 Nathaniel Craig posted the preprint “A gauge-invariant measure for lattice chiral fermions” on arXiv, proposing a non-perturbative construction of a gauge-invariant measure for overlap fermions on four-dimensional Euclidean lattices. The author presents the work as a candidate lattice definition for a class of non-abelian chiral gauge theories that includes the Standard Model. On 15 November OpenAI launched its science blog with the claim that the GPT-6 Astra model led the detailed proof. We were unable to read OpenAI's post directly: the model's role and Craig's quotes reach us through reports by AI Weekly and Digg, so the system's actual contribution should be treated as a claim by the developer, with no audit or independent verification.
On the purely theoretical side, arXiv lists the preprint at 60+63 pages and 12 figures. Zenodo hosts a mathematical compendium with a Lean formalization and a two-dimensional supplement with the computation logs. By comparing chiral projectors at successive lattice spacings, the formulation obtains an exponentially local current that satisfies Lüscher's reconstruction conditions. The construction, however, is stated to hold at finite lattice spacing and under a bound on field strength at the lattice scale. Central questions remain entirely open: reflection positivity at finite spacing, the definition of the interacting theory in continuum spacetime, and realistic simulations of the full Standard Model. Craig himself described the proposal as “far more an answer in principle than an answer in practice”.
According to reports on OpenAI's post, Craig said he set out the initial idea (“I had a strategy for tackling this problem”) and that “Astra helped turn that strategy into a construction and took the lead in developing the detailed proof”, showing how to extend the fields to finer lattices. Yet the arXiv abstract page makes no mention of assistance from AI systems; we have not checked the full text of the PDF. Craig names physicists Maarten Golterman, Andreas Kronfeld and Yigal Shamir among those who examined the proof and raised questions. The secondary sources disagree on what they concluded. According to AI Weekly, no independent assessment had been published as of 16 November, and we found no reactions from outside physicists and no articles in major science outlets.
The account of the collaboration comes from the author and is published by the model's developer; so far there has been neither peer review nor an independent audit. Without that scrutiny, and with no confirmation beyond company communications, the work remains an interesting proposal in principle, waiting for the theoretical physics community to test its foundations. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I opened the arXiv:2610.09093 abstract page and checked the title, single author, submission date (6 October 2026, v1), abstract, categories, page count, Lean formalization, Zenodo DOIs and ancillary files. OpenAI's post returned a 403; I confirmed its title and address via search. I read and compared the AI Weekly (16 November) and Digg reports: they agree on Astra's role, the limits and the “answer in principle” quote, and differ on the reviewers. I found no reactions from independent physicists. I ruled out that the topic had already been covered: the GPT-Rosalind-6 article concerns a different model and a different field.
- Incertezze
- I did not read OpenAI's post directly (it returns 403 to automated fetches); its existence, title and date (15 November) come from the search index and from the links in AI Weekly and Digg. GPT-6 Astra's role and Craig's quotes therefore come from secondary sources. The arXiv abstract page does not mention AI assistance; I did not read the full PDF. The account is Craig's, published by the model's developer: it is not an independent audit, and the preprint has not yet been peer-reviewed. The sources disagree on the three reviewers named: according to Digg the post reports no direct judgment from them, according to AI Weekly they examined the work privately. No public reactions from independent physicists and no coverage in major science outlets. The affiliation (UC Santa Barbara, OpenAI visiting scientist) does not appear on arXiv.
- Perché pubblicarla
- It is a concrete, verifiable case of an AI model contributing to a theoretical physics proof on a problem that has been open for decades. The preprint is public, with a Lean formalization, and the author himself states its limits. It lets us draw a precise line between “candidate” and “solved”, between a developer's account and independent verification: the site's anti-hype angle.