Open problems in 22 scientific fields already had solutions within reach, and an artificial intelligence system has found them. The credit goes to BootLoops, a toolkit developed by Harvard physicist Matthew Schwartz to use language models in precision quantitative science. In three months, the system contributed to preparing 36 manuscripts across 18 fields, with 19 coauthors, starting from about 400 candidate problems. The fields range from physics to linguistics, from ecology to health policy.
The project grew out of Schwartz's own experience. As Anthropic describes it, he stopped fighting Claude and began letting the model identify problems suited to its abilities, what he calls "Claude-shaped" problems. That led to BootLoops, which he describes as a harness for large language models doing exact quantitative calculations.
How the system works
BootLoops 1.0 is a scientific software package written or adapted to be driven by an LLM, paired with working protocols that make the model's results checkable. The GitHub repository lists twelve operational skills, and many of them are not research skills in the narrow sense: an acceptance gate, controls against known truths, and independence tracking that prevents any data source used to fit a model from certifying the result itself. Before touching real data, the system must recover a known answer on synthetic data.
Image: Anthropic
The principle is simple: the model can propose, but every result must pass checks a human researcher can retrace. On the project site, manuscripts are marked preliminary when they have not yet been fully verified by people, and many were first drafted by LLMs.
Verification still in progress
The question of reliability remains open. The project states that all manuscripts were read by humans for accuracy, but some have not yet been fully checked. A separate batch of AI-agent experiments on 40 health studies built on NHANES data found that in 36 cases the analysis behind the main result did not exactly match the one the paper described. That finding, however, has not passed review: two gpt-6 agents flagged major issues with it.
Image: bootloops.ai
The BootLoops case belongs to a series of experiments in which AI has surfaced overlooked connections. One group used 90 Claude agents to flag, in a 1999 chemistry paper, a compound potentially relevant to spintronic memory that had sat unnoticed for 27 years because no one read both fields at once. Here too, experimental confirmation is still pending.
The bottleneck moves
The real significance of the experiment may lie in where the problem has shifted. Finding connections across separate literatures is no longer the main obstacle: the harder task becomes verifying them. With 36 manuscripts in three months, the capacity to generate hypotheses is growing faster than the capacity to check them, and it is this gap that BootLoops' protocols try to close.
No comments yet. Be the first!