Harvard physics professor and AI institute principal investigator Matthew Schwartz explored whether modern artificial intelligence can perform theoretical physics research. Rather than testing fully autonomous end-to-end science systems, which he notes often struggle with nebulous physics problems, Schwartz supervised Claude through a concrete, second-year graduate-level calculation without touching a file himself. The specific problem focused on resumming the Sudakov shoulder in the C-parameter within quantum chromodynamics, a brutally difficult calculation with established principles but missing complete treatments. Initially, when Schwartz gave the research prompt to frontier models, they all failed. To pursue the experiment scientifically, he established strict rules and used a web interface to have multiple models merge their ideas into a master plan. Claude then broke this plan down into 102 separate tasks across seven stages, maintaining a tree of markdown files in VS Code using Claude Code. This organization allowed the AI to look up previous steps rather than trying to hold everything in a single long context window, demonstrating how structured supervision can guide models through complex scientific workflows where end-to-end autonomy currently falls short.
Vibe physics: The AI grad student
Harvard physics professor Matthew Schwartz tested whether an AI model could conduct a real theoretical physics research calculation from start to finish. He supervised Claude through a well-defined graduate-level problem involving quantum chromodynamics without touching a file himself. The experiment aimed to determine if large language models can handle intermediate scientific workflows.
Why it matters
This experiment highlights the current limitations and potential workflows for using large language models in complex scientific research, suggesting structured supervision and modular task breakdown are more effective than relying on fully autonomous end-to-end agents for theoretical physics.