hn.today

Claude-Shaped Science

anthropic.com28 points12 comments
Screenshot of Claude-Shaped Science

Matthew Schwartz describes a practical approach to resolving the "impedance mismatch" between what scientists want and what large language models actually do well. Rather than forcing a model to act like a human scientist, he built BootLoops, an open-source harness that exploits Claude's strengths - wide knowledge, fast parsing of papers and data, and strong coding and numerical skills - for a class of exact, checkable quantitative problems. He applied this to scattering-amplitude research using the semi-numerical S-matrix bootstrap: Claude Fable 5 ported and improved code, reproduced prior results in minutes that had taken weeks by hand, and generalized the machinery from logarithmic to elliptic integrals. In a few weeks BootLoops completed 30 integrals end-to-end, including 15 known reproductions and 15 previously uncomputed results, all verifiable by high-precision numerics.

Because many scientific problems share mathematical structure, BootLoops discoveries translated across disciplines. Techniques for Feynman integrals mapped to Bayesian-evidence calculations in phylogenetics, finite-field reduction methods found uses in evolutionary biology, and other Claude-shaped problems arose in ecology and population genetics. Many initial outputs were technically correct but required domain experts to steer them toward questions that matter to those fields. The broader claim is practical: align problems to current LLM capabilities, combine automated high-precision computation with expert guidance, and you can accelerate verifiable advances across diverse sciences.

Read on anthropic.com12 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in Science

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.