On October 1, 2026, Anthropic’s research site published “Claude-shaped science”, a guest essay by physicist Matthew Schwartz written with acknowledgments to 22 collaborators. Schwartz argues there is an impedance mismatch between what current models are good at and what working scientists ask of them, and that instead of making Claude imitate a human scientist one should look for “Claude-shaped problems”: quantitative work that rewards tirelessness, careful bookkeeping and code. He describes building a toolkit called BootLoops for this and reports a set of case studies.
The examples span fields. In physics, the work computed 30 Feynman integrals using semi-numerical bootstrap methods, 15 of them previously unsolved. In ecology, it found that tree species turnover on Barro Colorado Island happens 4.5 times faster than neutral theory allows. In population genetics, it analysed 5.7 billion mutation pairs from the 1000 Genomes Project and identified gene conversion mechanisms. In economics, it ported the replication packages of 4,452 papers, about 30,000 routines, from commercial to open-source software. In linguistics, it built AccStack, a database of word stress across 6,072 languages with 160,000 bibliography entries. Further work is mentioned in phylogenetics, cosmology, geochemistry, genomics and statistics.
Why it matters: it comes a week after Anthropic’s own posts on a Claude-led amplitude calculation and on Claude discovering a new reverse transcriptase family, and it offers a working theory of where AI helps science most right now: breadth of careful quantitative labour rather than a single eureka.
What it does not show: this is an essay with case studies, not a peer-reviewed paper, and it appears on the website of the company that makes the model. Schwartz himself stresses that technical correctness does not guarantee scientific significance, and that domain experts were essential for steering the work toward results that matter.