Vibe physics: The AI grad student

Source: Anthropic Research Date: 2026-03-23 URL: https://www.anthropic.com/research/vibe-physics

Summary

Professor Matthew Schwartz supervised Claude Opus 4.5 through a graduate-level theoretical physics calculation using text prompts only — 102 sequential tasks across seven stages, Claude Code for file management, cross-verification with GPT and Gemini. Result: arXiv paper on Sudakov shoulder resummation with novel factorization theorems and experimentally testable predictions, produced in two weeks vs. 1-2 years with a grad student. Expert oversight required to catch systematic errors; the human did no direct file editing.

Implications

This is the AI-accelerated scientific research thread with a credibility-establishing result — a real arXiv paper with novel content, not a demo. The “G2 level” (second-year grad student) framing gives a human-interpretable capability benchmark. The expert supervision requirement is the honest caveat: without the professor’s guidance on research direction and error correction, the result wouldn’t have been achievable. This is expert-in-the-loop research acceleration, not autonomous AI science. Watch for this as the template Anthropic uses to pitch scientific institutions, and for the two-week vs. 1-2 year compression ratio becoming a recurring talking point in science AI discussions.