Claude AI solves a year-long physics problem in just two weeks

Anthropic's Claude AI has completed a complex physics research task in just two weeks—a process that normally takes human researchers a full year. Harvard professor Matthew Schwartz guided the system through the work using only text prompts, marking a significant step in AI-assisted scientific research. Schwartz selected a well-defined physics problem suitable for a second-year graduate student: resumming the Sudakov shoulder in the C-parameter. Unlike earlier AI research tools that relied on brute-force trials, this project structured the work into over 110 separate drafts across seven stages. Each stage was documented in markdown files to improve Claude's ability to organise and refine its output.

The effort involved 36 million tokens and more than 40 hours of local CPU processing. Schwartz never directly edited files, instead using text prompts alone to steer the AI. While Claude demonstrated strong capabilities, its results still required careful review by a domain expert to ensure accuracy. Schwartz sees this as proof that large language models can now assist in theoretical research. The achievement follows a broader trend in 2024 and 2025, as companies like Sakana AI developed autonomous research systems to accelerate scientific discovery.

The project produced a technically rigorous paper in a fraction of the usual time. Schwartz's approach highlights AI's growing potential for symbolic manipulation in advanced research. Future work may build on these methods to tackle even more complex scientific challenges.