Artificial intelligence has achieved a breakthrough by solving complex problems that had remained insurmountable for scientists for years across 22 different scientific fields in just three months. Based on this work, 36 scientific preprints have already been published.
The key discoveries and solutions achieved by the system cover the following areas:
Physics and Mathematics:
Solved complex integral equations. Through applied patterns, the model uncovered deep mathematical connections that were previously not apparent to physicists.
Astrophysics and Historical Sciences:
Precise calculations were made regarding the dating of sunspots, as well as the decryption of an unknown medieval manuscript.
Biology and Evolution:
Accurately calculated the probabilities of phylogenetic (evolutionary) trees for humans, chimpanzees, gorillas, and orangutans.
Linguistics:
Created the massive ACCSTACK catalog, which analyzes and systematically classifies the patterns of accent and intonation in more than 6,000 languages worldwide.
Healthcare and Economics:
Calculated and optimized the rating thresholds for Medicare healthcare insurance plans, which is directly related to the redistribution of approximately $1 billion in bonuses.
How did the AI manage to find these undiscovered answers? The idea of solving a complex problem in one scientific field using methods from another is not new; for instance, Charles Darwin's theory of natural selection once inspired computer algorithms that find solutions by simulating evolution. However, human science is often "siloed": physicists do not read linguistics journals, and biologists may not master the mathematical apparatus of quantum mechanics.
As Georgia Chenning, Head of AI for Science at Hugging Face, notes, a physicist is unlikely to spend six months studying linguistics, whereas an AI can compare and "cross-pollinate" toolkits from different fields in seconds.
At the heart of these results is the BootLoops software tool, developed by Matthew Schwartz, a physicist at MIT's FutureTech and Harvard University, and a visiting researcher at Anthropic.
Back in December 2025, Schwartz had used Claude Opus 3.5 as a "virtual graduate student," preparing a physics research paper in two weeks, but later concluded that having the system work like a human researcher is inefficient.
BootLoops can work with any Large Language Model (LLM), including Anthropic's Claude and OpenAI's ChatGPT. Over three months, the system, with the help of Claude, identified problems in various fields that shared the same mathematical structure but were previously considered unsolvable, and provided their solutions.
The BootLoops experiment proves that the current advantage of AI is not in inventing new theories from scratch, but in finding connections between already accumulated knowledge that a specialist might have missed. The method may have existed for a long time, but it was applied only within one narrow direction.
At the same time, researchers emphasize that artificial intelligence lacks scientific "taste" and cannot independently assess whether a mathematically correct solution revolutionizes the understanding of a given field.
Given that the AI calculations in the Medicare example also concern public finances and the distribution of billions of dollars, where an error could have serious consequences, professional oversight and independent verification of the results remain indispensable.
