BootLoops Harness Helps Claude Solve Scientific Equations
Harvard physicist Matthew Schwartz has released BootLoops, an open-source harness that helps Anthropic's Claude model perform precise calculations to bridge gaps across scientific fields.
Developed by Schwartz and 19 co-authors, the BootLoops harness aims to target "Claude-shaped problems" that sit at the intersection of scientific inquiry and AI capability. Over a three-month period, the team utilized the open-source tool to produce 36 manuscripts across 18 different fields. In particle physics, Claude computed 30 elliptic integrals and scattering amplitudes; 15 of these reproduced existing scientific knowledge, while the other 15 were calculated for the very first time.
The tool's applications spanned multiple disciplines. In ecology, Claude solved a 20-year-old equation from neutral biodiversity theory, revealing that tree species composition on Barro Colorado Island is changing 4.5 times faster than previously theorized. In population genetics, the system analyzed 5.7 billion mutation pairs from the 1000 Genomes Project to find evidence of gene conversion. Additionally, the researchers used BootLoops to build an AI data editor that automatically checked 4,452 replication packages for economics journals, and compiled a word stress database covering 6,072 languages.
Despite these breakthroughs, Schwartz warns that the system is highly compute- and token-intensive, and requires constant human verification. Claude frequently suffers from a tendency to "declare victory too early" or brute-force mathematical calculations rather than finding elegant solutions. Schwartz noted that automated checks remain unreliable, and the model can easily draw incorrect conclusions even when its underlying calculations are entirely accurate.
For scientific practitioners, this development shifts the paradigm of research and education. Schwartz suggests that traditional training, such as introductory Python courses for engineers, is becoming obsolete because Claude can implement complex machine learning models on command. Instead of spending years on manual calculations, researchers must pivot to guiding AI systems, defining the right questions, and providing the essential domain expertise to verify automated outputs.
This is our own summary of reporting by The Decoder
