Anthropic’s guest post reports 30 completed integrals, a forest model and analysis of 5.7 billion mutation pairs using Claude-assisted research tools.
Anthropic published a guest post by professor Matthew Schwartz on October 1 describing BootLoops, his open-source toolkit for exact calculations in quantitative science. The system can work with different language models and is designed for research tasks in which mathematical methods, code and numerical results can be checked rigorously.
BootLoops emerged from Schwartz’s work on scattering amplitudes in theoretical physics. Rather than asking Claude to resolve deep conceptual questions, he used it to consolidate methods scattered across papers and programming languages, then create missing software. Schwartz said Claude reproduced a result from one of his papers in about 20 minutes, compared with the weeks he had spent writing the original code.
The toolkit was subsequently extended from logarithmic functions to more difficult elliptic integrals. Schwartz reported taking 30 integrals through the complete BootLoops process in a few weeks: 15 reproduced known results with the new method, while another 15 had not previously been computed, according to his account. Anthropic’s post does not present an independent assessment of the full set.
Claude also searched for problems in other disciplines with compatible mathematical structures, leading to applications in ecology and population genetics. Schwartz said the model’s initial findings were usually technically correct but often lacked scientific significance until subject specialists helped reformulate the questions.
In ecology, the researchers applied the computation to tree data from Barro Colorado Island in Panama and found that species composition changed 4.5 times faster than neutral theory permitted. After consulting plant biologist James O’Dwyer, they developed a model intended to separate random fluctuations from species-dependent effects. Schwartz said it agreed closely with the data and is now being extended to other forest plots.
For population genetics, the team examined 5.7 billion pairs of nearby mutations in genomes from the 1000 Genomes Project and reported evidence of gene conversion. Schwartz said many analyses based on linked genetic variation disregard this mechanism. He also described projects in phylogenetics, Earth science, genomics and sunspot research as still undergoing further exploration and verification.
Other projects moved beyond BootLoops-style exact calculations into broader work with Claude. An economics effort transferred replication packages from 4,452 papers out of commercial software and into open-source code. In linguistics, Claude and three specialists produced AccStack, a word-stress database covering 6,072 languages and accompanied by a bibliography of 160,000 phonology works. The post does not characterize either project as a direct BootLoops application.
The projects suggest a practical division of labor: a model can expand the search for methods, generate software and support checkable computation, while researchers determine whether a result is meaningful, challenge its assumptions and recruit relevant specialists. In Schwartz’s examples, technical correctness was not enough on its own to establish scientific value.
| Field | Role of Claude or BootLoops | Reported result or status |
|---|---|---|
| Mathematical physics | End-to-end BootLoops process for elliptic integrals | 30 integrals: 15 reproductions and 15 reportedly new computations |
| Ecology | BootLoops solved the neutral-theory equation; the model was refined with a domain expert | Species composition changed 4.5 times faster than the neutral prediction; the model is being extended to other forests |
| Population genetics | Exact methods and new BootLoops tools for mutation analysis | 5.7 billion pairs analyzed; evidence of gene conversion reported |
| Economics and linguistics | Broader work with Claude, not presented as a direct BootLoops application | Replication packages from 4,452 economics papers moved to open-source code; AccStack covers 6,072 languages |
Sources
Originally published by AI Feed: Anthropic Details BootLoops Research Across Physics and Biology.
Top comments (0)