Vibe Physics

Vibe physics: The AI grad student \ Anthropic Science Vibe physics: The AI grad student Mar 23, 2026

Can AI do theoretical physics? In this guest post, professor of physics Matthew Schwartz decided to find out by supervising Claude through a real research calculation, start to finish, without ever touching a file himself. His account of what happened is below. Summary I guided Claude Opus 4.5 through a real theoretical physics calculation, encapsulating the complexity of code and computations behind text prompts. The result was a technically rigorous, impactful high-energy theoretical physics paper in two weeks instead of the usual year. Over 110 separate drafts, 36M tokens, and 40+ hours of local CPU compute, Claude proved fast, indefatigable, and eager to please. Claude is impressively capable, but also sloppy enough that I found domain expertise essential for evaluating its accuracy. AI is not doing end-to-end science yet. But this project proves that I could create a set of prompts that can get Claude to do frontier science. This wasn’t true three months ago. This may be the most important paper I’ve ever written—not for the physics, but for the method. There is no going back.

Who am I? I’m Matthew Schwartz , a professor of physics at Harvard and a principal investigator in the NSF Institute for Artificial Intelligence and Fundamental Interactions ( IAIFI ). My area of expertise is quantum field theory, which asks what matter is, how particles interact, and why the Universe has the rules it does. One might say I wrote the book on the subject. I’ve been working with modern machine learning tools for over a decade. My first modern ML paper , from 2016, was an early application of deep learning to particle physics. In a Nature Reviews Physics piece in 2022, I compared the timescale of AI and human evolution, arguing that transferring understanding between biological and artificial intelligence would become a fundamental challenge. Since then, I’ve been trying to push AI towards more symbolic work (manipulating mathematical expressions rather than numerical data) and the core questions in theoretical physics. The hype There has been a lot of recent hype about AI scientists doing end-to-end research autonomously. In August 2024, Sakana AI released their AI Scientist , a system designed to automate the entire research lifecycle—from generating hypotheses to writing papers. In February 2025, Google released an AI co-scientist built on Gemini, promising to help researchers generate and evaluate hypotheses at scale. And in August 2025, the Allen Institute for AI (Ai2) launched the open-source Asta ecosystem, featuring tools like CodeScientist and AutoDiscovery to find patterns in complex datasets. Since then, a new entrant has appeared every few months—FutureHouse’s Kosmos , the Autoscience Institute’s Carl , the Simons Foundation’s Denario project, among others—each promising some version of end-to-end autonomous research. Even as these approaches are visionary, their successes to date seem a bit forced: run hundreds or thousands of trials and define the best one as interesting. While I believe we are not far from end-to-end science, I’m not convinced we can skip the intermediate steps. Maybe LLMs need to go to graduate school before advancing straight to the Ph.D. In mathematics, automated end-to-end AI agents have produced impressive results, at least for a certain class of problems. An early breakthrough was DeepMind’s FunSearch , launched in 2023, and later AlphaEvolve , which used LLMs to make new discoveries in combinatorics. A related project, AlphaProof , earned a silver medal at the 2024 International Mathematical Olympiad, solving problems that stumped all but five human contestants, and in 2025, an advanced version of Gemini achieved the gold-medal standard . And, just as in science, more achievements have continued to follow. What about theoretical physics? End-to-end AI scientists have found their footing in data-rich domains, but theoretical physics is not one of them. Unlike mathematics, theoretical physics problems can be more nebulous—less about formal proof search and more about physical intuition, choosing the right approximations, and navigating a landscape of subtleties that often trip up even experienced researchers. Even so, there are problems in physics where AI might be better suited. Not yet the paradigm-shifting questions at the frontier, but those where the conceptual framework is established and the goal well-defined. To find out if AI can solve these types of theory problems, I supervised Claude through a real research calculation at the level of a second-year grad student. Problem selection In grad school, at least at my institution, first-year theory students (G1s) typically just take classes. Research often begins in the second year. G2 students start with well-defined projects that have a guarantee of success—often follow-ups from previous studies where the methods are established and the endpoints clear. This gives them a chance to learn the techniques, make mistakes in a controlled setting, and build confidence. It's also easy for me as an advisor: I can check their work, spot where they've gone off track, and quickly reorient them. Advanced students (G3+) work on more open-ended, creative problems. These require choosing your own direction, deciding which approximations matter, and sometimes realizing the original question was wrong (such is the nature of research). For this experiment, I deliberately chose a G2-style problem. My reasoning was that LLMs can already do all the coursework, so they are past the G1 stage. But if AI can't do the G2 projects—the ones with training wheels, where I know the answer and can check every step—then it certainly can't do the G3+ projects where creativity and good judgment are essential. The problem I chose was resumming the Sudakov shoulder in the C-parameter. For context, when you smash electrons and positrons at a collider, debris sprays out; the C-parameter is a single number that describes the shape of that spray, and its distribution has been measured with extreme precision. The theory that's supposed to predict that distribution is quantum chromodynamics, the study of the strong nuclear force, which holds nuclei together and powers the sun. The C-parameter is well-defined on paper but brutally hard to calculate, so you approximate. Every approximation is a…