Open-source “BootLoops” framework helps AI models perform accurate scientific computations

Title: Anthropic Claude Powers BootLoops Across 18 Scientific Fields

Meta description: Anthropic Claude and BootLoops reshaped Harvard physicist Matthew Schwartz's science workflow, producing 36 manuscripts across 18 fields in three months.

Tags: Anthropic Claude, BootLoops, Matthew Schwartz, scientific AI, open-source harness

Featured image: Abstract technology-style network linking particle physics, ecology, genetics, economics, and linguistics; no real people, no company logos, no text.

Anthropic Claude has become more useful in scientific work when it is not asked to imitate a human researcher, according to Prof. Matthew Schwartz, a physicist at Harvard University who is also a visiting researcher at Anthropic. In an Anthropic guest post, Schwartz said he stopped treating Claude as a general researcher and began looking for "Claude-shaped problems," tasks that fit the strengths of today's AI models. That shift led to BootLoops, an open-source harness for exact scientific calculations whose source code is available on GitHub. The timeline shows how one problem-solving tool moved from particle physics into ecology, population genetics, economics, and linguistics, producing 36 manuscripts across 18 fields in three months with 19 co-authors.

Background: From Human Researcher to Claude-shaped Problems

Schwartz's starting point was a distinction between automated problem-solving and relevant research. A model can execute steps, but a scientist must decide which steps matter. "Claude-shaped problems" sit at the overlap between what scientists want to study and what AI can actually do.

He used the convex hull idea, a geometric concept describing the outer boundary enclosing a set of points, to frame human knowledge as fragmented fields with unexplored gaps. One lab can spend 20 years on a single set of genes using a single method, leaving nearby genes and other approaches unexamined. BootLoops is meant to bridge those gaps, linking uneven edges of established knowledge.

Development: BootLoops Moves From Physics to Cross-field Calculations

Schwartz began with particle physics calculations, specifically scattering amplitudes and elliptic integrals. In a matter of weeks, Claude produced 30 BootLoops integrals: 15 matched previously known results, and 15 were calculated for the first time.

The tool also cracked a 20-year-old equation in neutral biodiversity theory that had resisted large-scale computation. Using the data, it indicated that tree species composition on Barro Colorado Island in the Panama Canal is shifting 4.5 times faster than neutral biodiversity theory permits. Ecologist James O'Dwyer then guided the result into a stronger predictive model.

In population genetics, the team examined 5.7 billion mutation pairs from the 1000 Genomes Project and found support for gene conversion. Other outputs included an AI data editor for economics journals that automatically verified 4,452 replication packages, later published as an NBER Working Paper, plus a word stress database for 6,072 languages built with three linguists.

Latest: Speed, Reliability, and the Limits of Automated Research

The pace has changed how Schwartz thinks about scientific planning. He questions the logic of applying for a three-year grant for work that a model might complete overnight. The purpose of training PhD students is now uncertain. His earlier post on Vibe Physics already pointed this way, and the sense has strengthened since then.

For fields such as computer science, the disruption is especially striking. Two years ago he would have treated a Python for Engineers course as essential; now he sees it as unnecessary, since Claude can handle those tasks. Building ML models for physical phenomena is another area where AI can now take over. Deep knowledge of neural networks also yields smaller returns if Claude can implement current ML research on command.

Schwartz warns about the models' weaknesses. Claude often declares victory too early, and phrases like "done, with one asterisk" often mean "not done at all." It also misjudges task duration and often brute-forces calculations rather than finding elegant solutions. Automated checks are not reliable, and conclusions can be wrong even if calculations are correct. Beyond reliability, the model drifts toward old, heavily cited debates instead of genuinely new questions. The projects were "compute- and token-intensive," Schwartz says.

Current Stage and What Comes Next

The BootLoops timeline shows a present stage in which Anthropic Claude is not replacing scientific judgment but expanding where exact calculation and cross-field pattern search can happen. The results often became scientifically valuable only after domain experts stepped in to set direction.

He also sees risk in focusing on big math problems and headlines. Unrealistic expectations could pull attention away from productive applications already working today. Real progress will still be built on solid foundations, just faster. The scientific method is not threatened, and human guidance and taste remain indispensable. The next phase will likely depend on how labs decide which problems are worth automating and which still need human framing.