Matthew Schwartz details BootLoops toolkit for AI-assisted science

Anthropic official brand assets (anthropic.com)

Matthew Schwartz details BootLoops toolkit for AI-assisted science

The open-source project combines quantitative tools with Claude and expert guidance; Anthropic says Schwartz owns and maintains it.

Physicist Matthew Schwartz has described BootLoops, an open-source toolkit he built with Claude to automate exact calculations across scientific fields, in an October 1 guest post published by Anthropic.

Schwartz said the project grew from work on scattering amplitudes in mathematical physics. He then enlisted specialists to assess applications in ecology, population genetics, economics, linguistics and other disciplines.

Over three months, he said the effort produced 36 manuscripts across 18 fields with 19 coauthors, selected from about 400 candidate problems. Those figures and research results are Schwartz’s account, rather than an independent assessment.

Advertisement

In physics, he reported completing 30 integrals with the toolkit: 15 reproductions of known results using a new method and 15 calculations he said had not previously been completed.

Schwartz said an ecology project with plant biologist James O’Dwyer developed a predictive model after separating neutral fluctuations from species-dependent effects. A population-genetics collaboration with Michael Desai analyzed 5.7 billion pairs of nearby mutations from the 1000 Genomes Project and found evidence of gene conversion, according to the post.

Other projects included an AI data editor that converted replication packages from 4,452 economics papers into open-source code and checked results against published tables. An effort with linguists produced AccStack, a word-stress database covering 6,072 languages, Schwartz said.

He emphasized that technically correct calculations were often scientifically unremarkable until domain experts helped redirect the work. Additional findings across fields remain under further exploration and verification.

The workflow uses separate Claude Code sessions on Google Cloud virtual machines, with a coordinating session allocating compute and checking results. Schwartz said the work required substantial compute, tokens and human supervision, including checks on proofs, conclusions and the model’s tendency to declare success prematurely.

Anthropic disclosed that Schwartz worked as a visiting researcher at the company during the project. BootLoops is owned and maintained by Schwartz and is not an Anthropic project. Its harness is available on GitHub for others to use and contribute to.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.
Matthew Schwartz details BootLoops toolkit for AI-assisted science
Matthew Schwartz details BootLoops toolkit for AI-assisted science

The open-source project combines quantitative tools with Claude and expert guidance; Anthropic says Schwartz owns and maintains it.

Anthropic official brand assets (anthropic.com)

Physicist Matthew Schwartz has described BootLoops, an open-source toolkit he built with Claude to automate exact calculations across scientific fields, in an October 1 guest post published by Anthropic.

Schwartz said the project grew from work on scattering amplitudes in mathematical physics. He then enlisted specialists to assess applications in ecology, population genetics, economics, linguistics and other disciplines.

Over three months, he said the effort produced 36 manuscripts across 18 fields with 19 coauthors, selected from about 400 candidate problems. Those figures and research results are Schwartz’s account, rather than an independent assessment.

Advertisement

In physics, he reported completing 30 integrals with the toolkit: 15 reproductions of known results using a new method and 15 calculations he said had not previously been completed.

Schwartz said an ecology project with plant biologist James O’Dwyer developed a predictive model after separating neutral fluctuations from species-dependent effects. A population-genetics collaboration with Michael Desai analyzed 5.7 billion pairs of nearby mutations from the 1000 Genomes Project and found evidence of gene conversion, according to the post.

Other projects included an AI data editor that converted replication packages from 4,452 economics papers into open-source code and checked results against published tables. An effort with linguists produced AccStack, a word-stress database covering 6,072 languages, Schwartz said.

He emphasized that technically correct calculations were often scientifically unremarkable until domain experts helped redirect the work. Additional findings across fields remain under further exploration and verification.

The workflow uses separate Claude Code sessions on Google Cloud virtual machines, with a coordinating session allocating compute and checking results. Schwartz said the work required substantial compute, tokens and human supervision, including checks on proofs, conclusions and the model’s tendency to declare success prematurely.

Anthropic disclosed that Schwartz worked as a visiting researcher at the company during the project. BootLoops is owned and maintained by Schwartz and is not an Anthropic project. Its harness is available on GitHub for others to use and contribute to.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.