Open-source "BootLoops" harness helps AI models perform precise scientific calculations

Harvard physicist Matthew Schwartz built BootLoops, an open-source harness that helps AI models perform precise scientific calculations. It is described in a guest post on Anthropic's blog; Schwartz is a visiting researcher there. Source code is on GitHub.

Schwartz says he stopped using Claude like a human researcher and instead looked for "Claude-shaped problems," tasks that play to the strengths of today's AI models. The idea builds on the convex hull concept: human knowledge is fragmented, fields extend in different directions, and the gaps between them stay unexplored. BootLoops is meant to fill those gaps. The results often became valuable only when domain experts stepped in to set the direction.

36 manuscripts across 18 fields

Over three months, Schwartz and 19 co-authors produced 36 manuscripts across 18 fields. He started with particle physics: within weeks, Claude computed 30 integrals with BootLoops, fifteen reproducing known results and fifteen computed for the first time.

In ecology, the model solved a 20-year-old equation from neutral biodiversity theory that had been impossible to compute at scale. Applied to data from Barro Colorado Island in Panama, the results showed tree species composition changing 4.5 times faster than the theory allows; ecologist James O'Dwyer helped turn the finding into a better predictive model. In population genetics, the team analyzed 5.7 billion mutation pairs from the 1000 Genomes Project, finding evidence for gene conversion. Other projects include an AI data editor for economics journals that automatically checked 4,452 replication packages, and a word stress database covering 6,072 languages built with three linguists.

Limits and warnings

Schwartz warns that the models often declare victory too early: "done, with one asterisk," he says, frequently means "not done at all." They misjudge how long tasks will take, brute-force calculations instead of finding elegant solutions, and their automated checks are not reliable. Conclusions can be wrong even when the calculations are correct. They also gravitate toward old, heavily cited debates rather than new questions, and the projects proved compute- and token-intensive.

What it changes for science

Schwartz says AI is changing the pace of science so fast that planning ahead is becoming nearly impossible: a three-year grant makes little sense for a calculation a model might solve overnight, and training PhD students is an open question. Two years ago he called a "Python for Engineers" course essential; today it is unnecessary. Still, he argues, the scientific method is not threatened, and human guidance and taste remain indispensable.