Harvard theoretical physicist Matthew D. Schwartz has taken a practical route to putting artificial intelligence to work in the sciences. On October 1, 2026, he released BootLoops 1.0, an open-source toolkit built on one premise: AI performs best when it is given the right kind of problem.

Schwartz refers to these well-matched problems as “Claude-shaped” tasks. Over three months, he and his collaborators applied that framing to a pool of 400 candidate problems, ultimately producing 36 manuscripts with 19 coauthors.

What BootLoops Actually Does

BootLoops is a modular harness for scientific calculation. It sits between a researcher and a large language model, such as Anthropic’s Claude.

The toolkit’s flagship use case is mathematical physics. It can compute 30 integrals end-to-end. Half of those — 15 integrals — reproduce results that were already known. The other 15 are new, and they include elliptic Feynman integrals.

For non-physicists, Feynman integrals are the mathematics behind how particles interact in quantum field theory, and the elliptic kind is notoriously hard to evaluate.

Well Beyond Physics

The project does not stop at particle physics. BootLoops reaches into roughly 18 to 22 different fields, including ecology, population genetics, and cosmology.

Verification is central to the design. BootLoops produces its outputs as Python scripts, which return results to arbitrary precision. That choice addresses a known limitation of language models, which generate plausible text rather than exact calculations: by routing the numbers through executable code, BootLoops lets researchers check and re-run results instead of taking the model’s word for them.

How It Was Built

BootLoops grew out of an iterative process with Claude that began in December 2025 — roughly ten months before the public release. Schwartz started with Claude Opus 4.5. By the summer of 2026, the workflow had moved to Claude Fable 5, a mid-project change of underlying model.

The launch followed a guest post by Schwartz on Anthropic’s blog titled “Claude-shaped science,” which Anthropic published.

Despite its origins, BootLoops is model-agnostic. It is designed to work with other large language models, including Google’s Gemini and OpenAI’s ChatGPT, so research groups are not tied to a single AI vendor. The software ships under the MIT License, a permissive open-source license that allows anyone to inspect, use, modify, and redistribute the code.

Who Is Behind It

Schwartz is a Professor of Physics at Harvard. His work spans quantum field theory, particle physics, and machine learning, and he has also written a textbook on Quantum Field Theory.

The toolkit is explicitly framed as research instrumentation. It is not intended for clinical, actuarial, regulatory, or safety-critical decisions, keeping its scope squarely within research, where outputs can be independently checked.

What This Means for Research

The most useful idea in BootLoops may be the selection step rather than the software itself. Starting from 400 candidate problems and producing 36 manuscripts implies a considerable amount of filtering. Because the code is free to inspect and adapt, other labs can now test that selection logic on problems of their own. How the 36 manuscripts fare in peer review — and how widely other groups adopt a toolkit they are free to run — are the natural markers to watch next.