Anthropic Research
⭐⭐⭐⭐⭐
Vibe Physics: The AI Grad Student
Summary
Can AI do theoretical physics? A Harvard professor supervised Claude through a real research calculation—start to finish—without ever touching a file himself. The result: a technically rigorous, impactful high-energy theoretical physics paper in two weeks instead of the usual year.
Key Details
- 110+ separate drafts, 36M tokens, 40+ hours of local CPU compute
- Claude proved "fast, indefatigable, and eager to please"
- Domain expertise still essential for evaluating accuracy
- Method matters: "There is no going back"
Problem Selection
The problem chosen was resumming the Sudakov shoulder in the C-parameter—a second-year grad student (G2) level problem in quantum chromodynamics. The reasoning: LLMs can already do coursework (G1 level), but can they do the training-wheels projects where the advisor knows the answer?
The Method
- Only give text prompts to Claude Code—no editing files directly
- Encapsulated work into 102 separate tasks across seven stages
- Claude maintained a tree of markdown files—looking things up rather than remembering
- Task examples: "Task 1.1: Review BSZ Paper", "Task 1.2: Review Catani—Webber"
Key Findings
- AI is not doing end-to-end science yet
- But the author could "create a set of prompts that can get Claude to do frontier science"
- This wasn't true three months ago
- Maybe "LLMs need to go to graduate school before advancing straight to the Ph.D."
Core Insight: "This may be the most important paper I've ever written—not for the physics, but for the method."
Related AI Scientist Systems
- Sakana AI's AI Scientist (Aug 2024) - automates research lifecycle
- Google AI Co-Scientist (Feb 2025) - built on Gemini
- Ai2's Asta (Aug 2025) - open-source ecosystem with CodeScientist
- DeepMind's AlphaProof - silver medal at IMO 2024
- DeepMind's AlphaEvolve - discoveries in combinatorics