Vibe physics: An AI graduate student

Harvard physics professor Matthew Schwartz put Claude Opus 4.5 to work on an actual theoretical physics calculation, having it re-derive the Sudakov shoulder in the C-parameter distribution for electron-positron collisions at NLL level. The goal was to carry out a well-defined G2-type research task from start to finish using only text prompts, to see how far AI could go in conducting scientific research.

Rather than throwing everything at it at once, the work was broken down into 102 tasks, with a markdown summary left after each stage. The research flow proceeded in the order kinematics, NLO structure, SCET factorization, anomalous dimensions, resummation, matching, documentation, with each stage taking about 15 to 35 minutes.

There was plenty of trial and error along the way. Claude tried to move on to the next stage after finishing only half of the 14 tasks in Stage 1, and in Stage 2 it lost context and had to restart. In numerical calculations, it was strong at compiling EVENT2 and comparing histograms against analytical calculations, but it also frequently made basic mistakes like normalization and binning.

Even so, output accumulated quickly. 65 tasks were completed within three days, spanning everything from literature review to matrix element calculations in the soft/collinear limits, setting up SCET operators, and the first LaTeX draft. In the end, going through more than 110 individual drafts, 36 million tokens, and over 40 hours of CPU computation, the key point is that a theoretical physics paper that could once have taken about a year was completed in roughly 2 weeks.

The key conclusions are clear.

• AI still cannot fully carry out end-to-end science on its own.
• But with the right prompting and human domain verification, it can actually push through calculations at the level of frontier physics.
• The more important achievement, more than the physics result itself, is that it demonstrated the research workflow that makes this kind of work possible.