Harvard Physicist Introduces BootLoops, an AI Harness for Precise Scientific Calculations
Harvard physicist Matthew Schwartz has released BootLoops, an open-source AI harness that has already helped produce 36 scientific manuscripts across 18 fields.
A new approach to scientific research is emerging from the intersection of artificial intelligence and physical sciences. Professor Matthew Schwartz, a Harvard University physicist currently serving as a visiting researcher at Anthropic, has shifted how researchers interact with large language models. Rather than trying to force AI to think like a human researcher, Schwartz advocates for identifying "Claude-shaped problems"—tasks that align directly with the specific capabilities of modern neural networks. This methodology has culminated in the release of BootLoops, an open-source harness designed to facilitate precise scientific calculations, with its source code now publicly available on GitHub.
Schwartz uses the mathematical concept of a "convex hull" to describe the current state of human knowledge. While individual academic disciplines push outward in narrow, isolated directions, the spaces between these fields often remain completely unstudied. A research team might spend decades analyzing a specific set of genes using a single methodology, leaving adjacent possibilities unexplored. BootLoops is designed to bridge these gaps by identifying connections across the fragmented boundaries of various scientific domains.
Broad Scientific Achievements in Three Months
The practical application of the BootLoops harness has yielded rapid results. Over a three-month period, a collaborative effort involving 19 co-authors produced 36 manuscripts spanning 18 different academic fields. The initiative began with complex calculations in particle physics, focusing on scattering amplitudes and elliptic integrals. Utilizing BootLoops, Claude successfully calculated 30 highly complex integrals; 15 of these confirmed existing calculations, while the remaining 15 represented entirely new scientific findings.
The project quickly expanded beyond physics. In the field of ecology, the AI solved a 20-year-old equation originating from neutral biodiversity theory that had previously been considered too computationally demanding to solve at scale. When applied to real-world ecological data, the calculations revealed that the tree species composition on Barro Colorado Island in the Panama Canal is shifting 4.5 times faster than the theory predicts. This discovery allowed ecologist James O'Dwyer to assist in refining the predictive model.
Other notable successes spanned genetics, economics, and linguistics. In population genetics, the team processed 5.7 billion mutation pairs from the 1000 Genomes Project, uncovering evidence of a genetic mechanism known as gene conversion. For economics, the system functioned as an automated data editor, reviewing 4,452 replication packages for journals, which resulted in a published NBER Working Paper. Additionally, working alongside three linguists, the system constructed a word stress database covering 6,072 distinct languages.
Redefining Engineering and Academic Training
The rapid capability of these tools is forcing a reassessment of traditional academic and engineering pathways. Schwartz notes that the sheer speed of AI-driven calculations makes long-term planning incredibly difficult. Applying for a traditional three-year scientific grant to fund a calculation that an AI model might complete overnight is becoming increasingly impractical.
This shift has also raised questions about how to train PhD students and engineers. Schwartz points out that courses like "Python for Engineers," which he would have deemed essential just two years ago, are now largely unnecessary because models like Claude can handle these programming tasks directly. Similarly, building machine learning models to analyze physical phenomena is shifting, as AI can now implement state-of-the-art machine learning research on command, reducing the practical return on having deep, specialized knowledge of neural network architectures.
What it means for developers
For developers, the BootLoops project highlights both the immense potential and the significant engineering hurdles of using LLMs for complex, structured workflows. First, these advanced scientific workflows are highly compute- and token-intensive. Managing API costs and latency is a critical consideration for any team building similar analytical pipelines. To mitigate these expenses, developers can try top AI models cheaply through one API at https://apixoai.online, which provides a streamlined way to test and deploy multi-model architectures without managing multiple platform subscriptions.
Second, developers must build robust verification layers when designing systems that use LLMs for precise tasks. Schwartz warns that Claude frequently exhibits specific failure modes: it tends to declare victory prematurely—where a status of "done, with one asterisk" often translates to the task being incomplete—and it frequently miscalculates how long a process will take. Furthermore, the model often resorts to brute-forcing calculations rather than discovering elegant mathematical shortcuts, and its automated self-checks are not entirely reliable. Even when Claude's mathematical calculations are correct, its ultimate conclusions can still be flawed.
Finally, developers should note that AI models naturally gravitate toward well-documented, heavily cited historical debates rather than generating entirely novel inquiries. This means human guidance, domain expertise, and rigorous oversight remain essential components of any AI-driven development pipeline. The scientific method itself is not being replaced; rather, human "taste" and direction are the critical factors that turn raw AI outputs into valuable, validated results.
Source: Open-source "BootLoops" harness supports AI models in performing precise scientific calculations — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
