Anthropic publishes Matthew Schwartz's toolkit for AI-assisted science

Anthropic published an October 1st post describing BootLoops, an open-source toolkit built by Harvard physicist Matthew Schwartz that uses Claude to perform exact calculations across quantitative sciences, with Schwartz reporting 30 scattering-amplitude integrals completed — 15 reproductions of known results and 15 previously uncompleted calculations. Schwartz said the toolkit, available on GitHub and designed to work with different models, also supported an ecology analysis of forest biodiversity on Panama's Barro Colorado Island and a population-genetics analysis of 5.7 billion pairs of nearby mutations from the 1000 Genomes Project that found evidence of gene conversion. Schwartz cautioned that the figures come from his own account in the Anthropic guest post rather than an independent evaluation, and that models can produce technically correct calculations without identifying work scientists consider important.

Anthropic publishes Matthew Schwartz's toolkit for AI-assisted science The open-source BootLoops project targets exact calculations, while Schwartz says human experts still have to decide which results matter. By Ryan Merket https://runtimewire.com/author/ryan-merket · Published Primary source: X https://x.com/AnthropicAI/status/2105733864152858919 Why it matters BootLoops puts a reusable software layer around AI-assisted research, but Schwartz's examples show scientific judgment remains the scarce input: experts must choose questions and decide whether model-generated results matter. Anthropic's October 1st post https://www.anthropic.com/research/claude-shaped-science describes how Harvard physicist Matthew Schwartz https://mattschwartzphysics.net/ built BootLoops, an open-source toolkit intended to help AI models perform exact calculations across quantitative sciences. Schwartz says Claude used it to connect techniques from mathematical physics with problems in ecology, genetics and other fields. His account also makes the boundary of the claim clear: models can produce technically correct calculations without identifying work that scientists consider important. BootLoops is Schwartz's project, not an Anthropic product. The professor, who works on particle physics and machine learning, says he assembled the toolkit from software and protocols developed while using Claude on research calculations. It is designed to be used with different models, according to his post. The project is available on GitHub https://github.com/BootLoops-ai/bootloops/tree/66b680ce742e654cfe86da4f072a69061fe182b1 , and Schwartz says it combines computational tools with workflows for carrying out scientific projects. The approach grew out of a problem Schwartz calls an "impedance mismatch": scientists often approach models as human collaborators, while current systems are better at bounded tasks such as coding, parsing papers and performing structured calculations. In his account, BootLoops tries to present work in a form that fits those strengths. The bet is on a specialized workflow around the model, rather than asking a general-purpose chatbot to choose and complete an entire research agenda. Schwartz says the toolkit first helped Claude reproduce existing calculations in mathematical physics and extend methods to additional problems. He reports 30 scattering-amplitude integrals completed using the workflow: 15 reproductions of known results and 15 calculations he says had not previously been completed. Those figures describe Schwartz's account in Anthropic's guest post; they are not, by themselves, an independent evaluation of the toolkit's performance. He then followed mathematical connections into other disciplines. In ecology, Schwartz and plant biologist James O'Dwyer used BootLoops to analyze a model of forest biodiversity. Schwartz says their first result showed that tree-species composition at Panama's Barro Colorado Island changed faster than a neutral-theory model predicted. O'Dwyer, he writes, thought that finding alone would draw little interest from ecologists and helped redirect the work toward a model that could better characterize species' life histories. Schwartz says the resulting model matched data and is being extended to other forest plots. In population genetics, Schwartz reports that the researchers analyzed 5.7 billion pairs of nearby mutations from the 1000 Genomes Project and found evidence of gene conversion, a process he says is commonly left out of analyses using linked genetic variation. The post describes this and several other projects as ongoing work with collaborators. It does not present a single completed, independently validated body of research; Schwartz says additional projects are undergoing further exploration and verification. That qualification is central to the method. Schwartz writes that Claude's cross-field suggestions were often technically sound but scientifically unremarkable until researchers with relevant expertise helped shape the question. In the forest project, O'Dwyer's judgment changed what the team studied. In genetics, collaborator Michael Desai helped direct attention from an initial calculation toward relationships between mutation pairs. The human contribution was choosing what to pursue and assessing whether an answer mattered, not simply checking the model's arithmetic. Schwartz's earlier account of using Claude in theoretical physics https://www.anthropic.com/research/vibe-physics?vid=87 , published in March 2026, focused on a supervised calculation in his own field. BootLoops extends that experiment into a reusable, cross-disciplinary workflow. Anthropic's post says Schwartz coordinated multiple projects through separate Claude Code sessions, with agents running calculations and intermediate work organized in files. Schwartz also describes failures: Claude could declare a task finished before resolving a central problem, misjudge how long work would take, and pursue long calculations where building a new tool would have been more effective. The toolkit's promise, then, is narrower and more concrete than autonomous science. It can help automate technical work on problems that can be expressed and checked, while researchers supply field knowledge, scrutiny and direction. Schwartz argues that this division could help scientists work across disciplinary boundaries. His own examples show the unresolved cost: a model can generate candidate findings quickly, but human experts still have to separate a useful result from an impressive calculation with little scientific consequence.