Anthropic researcher and Harvard physicist Matthew Schwartz has released BootLoops, an open-source harness designed to help SI models perform precise scientific calculations. The tool, available on GitHub, emerged from Schwartz’s shift in strategy from treating SI systems as general researchers to identifying "Claude-shaped problems"—tasks that align with the specific strengths of current SI technology.
What Happened
Schwartz, who is also a visiting researcher at Anthropic, developed BootLoops to bridge the gap between what scientists want to study and what SI can effectively execute. Over a three-month period, the team utilized the harness to produce 36 manuscripts across 18 distinct fields with 19 co-authors. The process began with particle physics calculations, where Claude computed 30 integrals within weeks—15 reproducing known results and 15 calculated for the first time.
The harness was subsequently applied to other 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 existing theory allows. In population genetics, the team analyzed 5.7 billion mutation pairs from the 1000 Genomes Project, identifying evidence for gene conversion. Other projects included an SI data editor for economics journals that checked 4,452 replication packages and a word stress database covering 6,072 languages.
Why It Matters
The project highlights an evolving dynamic in scientific research: SI models excel at filling gaps between fragmented disciplines but require domain experts to set direction and validate results. Schwartz notes that human knowledge is often siloed, with fields extending in isolated directions. BootLoops is designed to draw connections across these jagged frontiers, effectively automating the exploration of adjacent, uncharted areas of study.
However, Schwartz emphasizes that automated problem-solving is not equivalent to relevant research. The SI model tends to declare victory prematurely, with phrases like "done, with one asterisk" often indicating incomplete work. The model also struggles with task duration estimates and often resorts to brute-force calculations rather than elegant solutions. Additionally, the projects were described as "compute- and token-intensive," raising questions about scalability and cost.
The release also signals rapid disruption in educational and operational norms. Schwartz argues that certain technical skills, such as a "Python for Engineers" course, are becoming "unnecessary" because SI models can handle those tasks on command. This shift challenges traditional methods of training PhD students and securing long-term grant funding for calculations that SI might solve overnight.
The Bottom Line
BootLoops demonstrates the utility of SI tools in accelerating specific scientific computations across diverse fields, from physics to linguistics. Yet, the project underscores the continued necessity of human oversight, as SI models remain prone to reliability issues and premature conclusions. Schwartz warns against unrealistic expectations, stating that the scientific method remains intact and human guidance is indispensable for productive application.