Harvard Physicist Uses Claude to Produce Frontier Physics Paper in Two Weeks
Summary
- • Harvard physics professor produced a frontier paper in two weeks using Claude Opus 4.5.
- • Project required 110+ drafts, 36M tokens, and 40+ hours of local CPU compute.
- • Claude was fast and capable but sloppy enough that domain expertise was essential.
- • Author calls it a turning point in AI-assisted science — 'there is no going back.'
Details
Frontier physics paper produced in two weeks with Claude Opus 4.5
Schwartz describes the result as a technically rigorous, impactful high-energy theoretical physics paper. The process involved 110 separate drafts and 36 million tokens, compressed into roughly two weeks versus the typical timeline of about a year.
110 drafts, 36M tokens, 40+ hours of local CPU compute
These figures give a concrete sense of the iteration volume and computational cost involved in using Claude as a research collaborator on a single physics paper — intensive and iterative, not a simple prompt-and-publish pipeline.
Domain expertise was essential — Claude was fast but sloppy
Schwartz stresses that while Claude was fast and eager to please, it made enough errors that a domain expert was necessary to evaluate its outputs. He explicitly states AI is not doing end-to-end science yet, and the human's role shifted from doing the work to supervising and validating it.
Author calls this the most important paper he has written — for the method, not the physics
Schwartz frames the significance not in the scientific findings but in demonstrating a repeatable method for using LLMs to contribute to frontier science — something he says was not possible three months earlier. He concludes 'there is no going back.'
Existing AI scientist projects criticized for selecting best result of many automated trials
Schwartz references Sakana AI's AI Scientist (August 2024), Google's AI co-scientist on Gemini (February 2025), and AI2's Asta ecosystem including CodeScientist and AutoDiscovery (August 2025), arguing their successes seem forced rather than genuinely autonomous.
AI has had more documented success in mathematics than in theoretical physics
DeepMind's FunSearch (2023) and AlphaEvolve made new discoveries in combinatorics. AlphaProof earned a silver medal at the 2024 International Mathematical Olympiad, and in 2025 an advanced version of Gemini achieved the gold-medal standard at the IMO.
Author suggests LLMs may need an intermediate phase before fully autonomous science
Schwartz argues the field should not skip intermediate steps on the path to end-to-end AI science, using the analogy that LLMs may need to go through a graduate school phase before reaching full scientific autonomy.
Research = scientific work and findings, Stat = quantitative data points, Insight = author's argued position or analysis, Context = background on related projects and systems, Industry Update = broader field developments
What This Means
A Harvard physics professor has demonstrated that a skilled domain expert working closely with Claude Opus 4.5 can compress roughly a year of theoretical physics research into two weeks — but only because the expert caught the model's frequent errors. This is not a story about AI replacing scientists; it is about AI dramatically accelerating the work of experts who know enough to supervise it. The author's view is that a method now exists for using LLMs to do frontier science, and that this represents a qualitative shift from what was possible just months ago.
Sources
- Vibe physics: The AI grad studentAnthropic