A Harvard physicist describes how he stopped asking a model to think like a scientist and started asking it to do the math. Matthew Schwartz, a visiting researcher at Anthropic, went looking for what he calls Claude-shaped problems, and the hunt produced a tool.
Schwartz named it BootLoops. The harness, released as open source, is aimed at exact scientific computation. His starting premise is fragmentation: human knowledge sits in disconnected pockets. One gene set can absorb two decades of a lab’s attention while its neighbors go untouched, and the tool exists to span those gaps.
Early output was concrete. Particle physics came first: 30 integrals worked through in weeks, half of them reproducing known results and half new. Ecology followed, where an equation from neutral biodiversity theory had resisted computation for years.
Applied to field data from Barro Colorado Island, that equation showed tree composition shifting 4.5 times faster than the theory permits. A separate genetics effort scanned 5.7 billion mutation pairs and surfaced signs of gene conversion. Across three months the work produced 36 manuscripts spanning 18 fields.
Schwartz does not oversell it. The model declares victory early, misjudges how long tasks take, and brute-forces where an elegant route exists. Human judgment, he says, is still what turns output into research.
