A physicist working at Anthropic reports 36 manuscripts in three months by matching Claude to “Claude-shaped” problems

Anthropic published an essay on 1 October by Matthew Schwartz, a physicist who has been working as a visiting researcher at Anthropic. His argument is that language models struggle in science less because they lack ability than because scientists ask them the wrong kind of question.

He calls the right kind “Claude-shaped”: “Many fields have problems that a technique from mathematics, physics, or computer science would solve outright if anyone knew it existed.” A model that knows many fields' methods can carry a technique across to a field that has never used it.

His tally is “36 manuscripts in 18 fields with 19 coauthors over three months, out of some 400 candidate problems”. In physics, he writes: “Soon we had 30 integrals BootLooped from end to end, comprising 15 reproductions of known results by this new method and 15 that had never before been computed.”

The method is released as BootLoops, an open-source harness on GitHub. The page says it is not an Anthropic project; Schwartz owns and maintains it.

These are the author's own claims, on his employer's site. The essay does not say how many of the 36 manuscripts have passed peer review, and no outside check was found in the window. It follows Anthropic's 24 September claim about a phage enzyme system, which outside biologists then narrowed, so the publication record of these papers is the thing to watch.

• Claimed 36 manuscripts in 18 fields with 19 coauthors over three months, out of some 400 candidate problems. Anthropic (Matthew Schwartz)
• Claimed 30 integrals computed end to end with BootLoops: 15 reproductions of known results and 15 never computed before. Anthropic (Matthew Schwartz)
• Confirmed The BootLoops harness is on GitHub; the page says it is owned and maintained by Schwartz, not Anthropic. Anthropic (Matthew Schwartz)