Visiting Researcher Matthew Schwartz Details 'Claude-Shaped Science' and Open-Source BootLoops Toolkit

Prof. Matthew Schwartz, a visiting researcher at Anthropic, proposes the concept of Claude-shaped science, arguing that LLMs excel when applied to problems matching their strengths—such as coding, mathematics, and rapid data processing—rather than attempting to mimic human scientific intuition directly. To facilitate this, Schwartz developed BootLoops, an open-source harness designed to enable precise quantitative calculations for LLMs. Schwartz emphasizes that BootLoops is not an Anthropic project but is owned and maintained by him.

High-Energy Physics and Ecology Applications

In high-energy physics, the BootLoops harness, powered by Claude Fable 5, utilized the S-matrix bootstrap method to reproduce and generalize scattering amplitude calculations. The system completed 30 integrals, comprising 15 reproductions of existing results and 15 first-time calculations, extending from logarithmic function series to elliptic functions. In ecology, BootLoops solved equations from Hubbell’s neutral biodiversity theory that had remained computationally intractable for 20 years. Analysis of data from Barro Colorado Island revealed that tree species composition changes occur 4.5 times faster than neutral theory allows, suggesting significant roles for natural selection and competition.

Population Genetics and Cross-Disciplinary Applications

In population genetics, the toolkit analyzed 5.7 billion adjacent mutation pairs from the 1000 Genomes Project. This analysis provided evidence for gene conversion mechanisms, a factor previously ignored in population history and disease mapping studies. The project also expanded into economics, porting replication packages for 4,452 papers from five major journals into open-source code, and linguistics, building the AccStack database with word stress information for 6,072 languages.

Operational Scale and Limitations

Over three months, the BootLoops framework supported 18 fields, 19 coauthors, and 36 manuscripts, selecting from approximately 400 candidate problems. The infrastructure ran on Google Cloud VMs with Claude Code sessions, utilizing GitHub and Overleaf integrations. Schwartz notes that while AI is powerful for finding hidden patterns in vast datasets, it lacks a sense of time, tends to overstate completion, and requires domain experts to redefine problems for scientific impact. The project emphasizes that AI should fill the "convex hull" of human knowledge—areas reachable but unexplored—rather than replacing the human-led scientific loop.