Physics legend Edward Witten credits Claude with key ideas in new paper

What's going on

Edward Witten is one of the most influential mathematical physicists in the world. He's still the only physicist ever to win the Fields Medal, the highest honor in mathematics. Now he has publicly named an AI as a contributor. In a new preprint on arXiv, the big online archive for research papers that haven't yet been peer-reviewed, he thanks Anthropic's Claude Opus 5.5. The model helped him with specific technical steps, he says. The news was reported by hyper.ai.

Witten isn't alone. In recent weeks, two other physicists have published papers on the same physics problem that rely heavily on AI: Nathaniel Craig with OpenAI's GPT-6 Astra, and Stephen D. H. Hsu with a mix of GPT and Claude models. Taken together, the three cases show that language models now do far more in theoretical physics than summarize the literature or write simple code.

The problem: chiral fermions on the lattice

All three papers deal with a long-standing obstacle in particle physics. To simulate theories like the strong nuclear force on a computer, physicists replace continuous space with a lattice, a regular grid of points, a bit like a photo made of pixels. For gauge theories, the mathematical descriptions of the fundamental forces, this works very well in many cases.

Things get tricky with chiral fermions. These are matter particles whose "left-handed" and "right-handed" versions behave differently. You need exactly these particles for the Standard Model of particle physics, for example to describe the weak nuclear force. The Nielsen-Ninomiya theorem says that under the usual assumptions, you can't cleanly put such a particle on a lattice. Unwanted extra particles show up, an effect known as fermion doubling. Physicists have been looking for ways around this for decades.

Witten's approach and Claude's contribution

Witten studies a framework in which the gauge fields are first discretized on a lattice, meaning they're broken down into individual values. They're then extended back into continuous space so the chiral fermions can be built in consistently there. The lattice acts as scaffolding, while the fermions live in smooth space.

According to hyper.ai, Witten describes two concrete contributions from Claude Opus 5.5 in the technical discussion:

• The model suggested an adjustment to the coordinate map. This lets the extended gauge fields pass smoothly from one lattice cell to the next while still matching the original lattice data.
• Claude also proposed replacing a more complicated step-by-step construction with a linear interpolation, a simple, even blend between known values.

Witten credits Claude in both the main text and the acknowledgments. He does add a caveat, though: his method is computable in principle, but not yet practical with standard Monte Carlo methods (simulations based on many random samples, the go-to tool in lattice physics).

This isn't the first time. Back in June, Witten published a paper on "entanglement wedge reconstruction," a topic from holography, which links quantum gravity to quantum theories living on a kind of boundary surface. In a footnote there, he thanked Claude Opus 4.8 for generalizing a calculation.

Two more attacks on the same problem

On October 6, physicist Nathaniel Craig published a paper whose result, in his own words, couldn't have happened without OpenAI's GPT-6 Astra. The core idea was his. He then worked on it over many rounds with the model and with human experts. GPT-6 Astra supplied numerous mathematical derivations, which Craig revised and organized. The team checked parts of the derivation with Lean, software that verifies mathematical proofs step by step for errors. The manuscript runs to more than 100 pages.

A few days later, Stephen D. H. Hsu released his own paper using a hybrid lattice-continuum approach. He said he had used GPT and Claude extensively. GPT handled much of the detailed work in the appendices, while Claude and other models helped with verification. Hsu stressed that he remains responsible for the results. He later said Witten had shown him his draft. The two had pursued similar approaches, but with different AI workflows. Hsu summed up his experience like this: GPT is stronger than Claude at creative, generative reasoning. Weaker models can still be useful for spotting errors, which stronger models then evaluate.

How the field reacted

Mathematician and physics blogger Peter Woit, usually known as a critical observer, called it encouraging that AI can help a top researcher like Witten. At the same time, according to hyper.ai, he cautioned that progress remains incremental and depends on humans checking the results carefully. On physics blogs and online forums, the discussion focused on two sides of the same coin. On one hand, the models open up access to complex areas of mathematics. On the other, someone has to check those long, AI-assisted proofs. Many point to Craig's 100-plus-page manuscript as an example.

Our take

When someone like Edward Witten credits an AI in the main text of a research paper, that carries weight. If he says a model came up with a useful technical idea, physicists take it seriously. The three papers point to a shift: large language models are becoming tools for real theoretical work, from derivations to trying out proof strategies to checking calculations.

Still, you shouldn't read too much into these cases. All three papers are preprints that haven't finished peer review. According to the authors, the core ideas came from the humans themselves, while the AI supplied building blocks and computational legwork. Witten's method isn't practical yet. And the more text the models generate, the more work goes into checking it. That moves the bottleneck rather than removing it. Tools like Lean could help here. The competition behind it all is interesting too: OpenAI and Anthropic are racing to see whose model is most useful for cutting-edge research. But Hsu's view that GPT is generatively stronger reflects one researcher's experience, not a systematic comparison.