‘Science in Claude’s style’: physicist Matthew Schwartz has written 36 papers in three months

Rather than having the model play the role of a researcher, the Harvard professor sets it specific mathematical and computational problems: elliptic integrals, ecology and genetics.

In a text published on October 1 by Anthropic, Harvard physicist Matthew Schwartz describes an "impedance gap" between what researchers expect from AI and what it really does. His advice: not to ask the model to imitate a human scientist, but to identify the problems he has created — exact, very computational, on a multi-disciplinary basis.

BootLoops and its first results

Schwartz has developed BootLoops, an open source toolbox that makes Claude iterate on verifiable quantitative calculations. He drew 30 resolute elliptic integrals, 15 of them never published, and 36 manuscripts in 18 domains in three months, with 19 sponsors. Examples cited include an equation of the neutral theory of biodiversity that has remained open since 2005, the analysis of 5.7 billion pairs of mutations to highlight gene conversion, the translation of 4,452 economics articles into some 30,000 code routines, or a linguistic basis covering 6,072 languages.

The physicist insists on the method: the advances come from going back and forth verified on real data, not from a single request. AI fills the gaps in highly fragmented human knowledge and frees researchers from tedious calculations to focus on conceptual orientation.