Physicist releases BootLoops, an open toolkit for exact LLM calculations
Harvard's Matthew Schwartz says BootLoops, a free harness for any language model, helped finish 30 Feynman integrals, 15 of them calculations not previously completed.
Matthew Schwartz, a Harvard physicist, published BootLoops on Thursday, a toolkit for exact calculations in quantitative science. Anthropic ran his account on its Science Blog. Schwartz owns and maintains the project, not Anthropic, though he was a visiting researcher there while building it. The code is open source on GitHub and works with any language model.
https://x.com/AnthropicAI/status/2105733864152858919
Schwartz argues there is an "impedance mismatch" between LLMs and science. In his words, "Claude and GPT are good at science, but they are not scientists." Treating a model as a human collaborator is not the best way to draw out its strengths. BootLoops instead works as a harness around the model.
His post reports 30 elliptic Feynman integrals completed with the workflow. Fifteen reproduced known results, and he says the other 15 had not previously been calculated. Schwartz says the project covered 36 manuscripts in 18 fields with 19 coauthors over three months. These figures come from Schwartz. We have not checked the physics.
The post says similar calculations recur in distant fields, and that Claude found links to ecology and population genetics. It cites an analysis of 5.7 billion pairs of mutations from the 1000 Genomes Project, and 4,452 economics papers ported to open code. About 400 candidate problems were explored in total.
No independent review of the results has been published that we found. Schwartz's own account describes the claims as his and his collaborators', and the post makes no suggestion that models can do original science unaided. The repository and a project site at bootloops.ai are open for anyone to try.