Harvard Physicist Uses Open-Source BootLoops and Claude to Draft 36 Papers in Three Months

Harvard physicist Matthew Schwartz has spent the past three months testing how far large language models can go in producing publishable scientific work. Using an open-source harness called BootLoops alongside Anthropic's Claude, he generated 36 manuscripts spanning 18 disciplines, from particle physics to linguistics. The volume alone is unusual, but the more important finding is what happened after the drafts existed.

BootLoops is not a model itself. It is a scaffolding tool that lets a general-purpose AI system carry out precise calculations and structured reasoning steps that raw chatbots typically fumble. Schwartz, a theoretical physicist known for work on quantum field theory, built the harness to force the model through defined computational stages rather than letting it guess at numbers. That distinction matters because scientific claims live or die on exact arithmetic and reproducible derivations.

The 36 manuscripts covered a wide range: particle physics problems, statistical analysis, and even topics in linguistics. In several cases, Schwartz reported, the AI-generated drafts contained errors that only became apparent when a domain expert reviewed them. The manuscripts were not ready for submission as they stood. Their value emerged when human specialists stepped in to check equations, catch false assumptions, and rewrite sections that sounded plausible but were wrong.

"Look at everything yourself," Schwartz advised, according to his account of the project. That warning cuts against the usual pitch for AI research assistants, which tend to emphasize speed and autonomy. Schwartz's experience points to a narrower role: the model as a fast first-drafter and calculation aid, with humans still responsible for verification. The gap between a fluent manuscript and a correct one remains wide.

The project lands as scientific publishers and universities debate how to handle AI-assisted submissions. Journals including Nature and Science have tightened disclosure rules, and many now require authors to state exactly which tools were used and how. Schwartz's output, produced in a single quarter, would likely require extensive documentation under those policies. It also raises a practical question for peer review: how does a referee evaluate a paper when the first draft came from a model that cannot explain its own reasoning?

BootLoops is open source, which means other researchers can replicate or modify the approach. That could accelerate similar experiments in chemistry, biology, and economics, where calculation-heavy work is common. It could also produce a flood of plausible-looking but flawed preprints if users skip the human verification step Schwartz insists on. The tool lowers the barrier to generating scientific text; it does not lower the barrier to getting the science right.

What to watch next is whether any of Schwartz's 36 manuscripts survive formal peer review, and whether collaborators outside Harvard adopt BootLoops for their own workflows. The experiment's real test is not how many drafts a model can produce in three months but how many of those drafts become correct, citable contributions. Schwartz's own advice, to check everything personally, may end up being the most durable takeaway.