Researcher builds a toolkit for Claude-shaped science

Harvard physicist Matthew Schwartz described BootLoops, a toolkit he developed after changing how he used Claude for scientific work. Rather than asking the model to imitate the way a physicist normally approaches a problem, he looked for tasks suited to the model's ability to manipulate and compare exact calculations.

The approach surfaced mathematical connections across fields including ecology and population genetics. Schwartz found that many of those links were technically correct but scientifically unremarkable, so he worked with domain experts to identify which questions were worth pursuing.

The distinction is useful for AI-assisted research: finding a valid connection and finding an important one are separate tasks. Automating more of the first does not remove the need for subject-matter judgement in the second.