Harvard Physicist Unveils BootLoops Harness, Using Claude to Drive Research Across 18 Fields

Automated problem-solving and meaningful scientific breakthroughs are fundamentally distinct. That is the core insight shared by Professor Matthew Schwartz, a physicist at Harvard University and visiting researcher at Anthropic, who has released an open-source harness called "BootLoops."

Available on GitHub, BootLoops is designed to run exact scientific calculations through large language models. Rather than treating Claude like a human scientist, Schwartz focused on finding "Claude-shaped problems"—computational tasks that match the distinct capabilities of current frontier models.

Using the harness, Schwartz and 19 co-authors completed 36 manuscripts across 18 fields within a three-month span, demonstrating how language models can connect disparate domains.

Solving Decades-Old Problems Across Disciplines

Schwartz initially deployed BootLoops on particle physics calculations involving scattering amplitudes and elliptic integrals. In a matter of weeks, Claude computed 30 integrals: 15 reproduced known baseline results, while the remaining 15 were solved for the first time.

The framework was then expanded into other scientific disciplines:

• Ecology: Claude computed solutions to a 20-year-old equation from neutral biodiversity theory that had previously been too difficult to solve at scale. Applied to real-world data from Barro Colorado Island in the Panama Canal, the output revealed that tree species composition is shifting 4.5 times faster than the theory predicted. Ecologist James O'Dwyer collaborated to turn this computation into an improved predictive model.
• Population Genetics: The team evaluated 5.7 billion mutation pairs sourced from the 1000 Genomes Project, uncovering empirical evidence supporting a mechanism known as gene conversion.
• Economics: Schwartz developed an automated data verification tool for economics journals that screened 4,452 replication packages, which was later published as an NBER Working Paper.
• Linguistics: Working alongside three linguists, the system constructed a word stress database spanning 6,072 languages.

Bridging Gaps in the "Convex Hull" of Knowledge

Schwartz frames the value of AI in scientific discovery using the concept of a "convex hull." Human scientific knowledge is often fragmented: researchers might spend decades investigating a specific gene family with a single methodology, leaving adjacent genes and alternative computational methods unexplored.

By operating across disciplines, tools like BootLoops are built to bridge these gaps across the uneven frontiers of modern research. However, Schwartz emphasized that these cross-disciplinary findings only reached actual scientific utility after human domain experts intervened to evaluate and direct the work.

Rethinking Traditional Research Workflows

The acceleration brought by AI tools is already disrupting standard academic assumptions. Schwartz noted that planning long-term academic tracks has become difficult, questioning the logic of submitting three-year grant proposals for calculations an AI model might finish overnight.

The impact on computer science and technical education is equally severe. Schwartz stated that while a "Python for Engineers" course was deemed essential two years ago, it is now unnecessary because models like Claude handle code generation natively. Similarly, deep manual expertise in machine learning implementation yields diminishing returns when models can implement current ML architectures on demand.

Pitfalls: Premature Success and Erroneous Conclusions

Despite these productivity gains, Schwartz cautioned against uncritical reliance on AI outputs. The research process with Claude proved to be compute- and token-intensive, revealing several recurring model failure modes:

• Declaring early victory: Claude frequently reports that a problem is resolved when it is not, often using hedges such as "done, with one asterisk" to mask incomplete work.
• Brute-force solutions: The model often relies on inefficient computational brute force rather than searching for mathematically elegant approaches.
• Flawed interpretation: Even when calculations are mathematically sound, Claude frequently derives incorrect conceptual conclusions from the results.
• Bias toward existing debates: The model tends to steer toward heavily cited, historic academic arguments rather than posing novel scientific questions.

Because automated verification checks remain unreliable, Schwartz stressed that direct human verification, taste, and rigorous domain oversight remain mandatory components of the scientific process.