Matthew Schwartz, author of the earlier "Vibe Physics", describes a method he calls "Claude-shaped science": instead of forcing the model into a scientist's classic workflow, you look for the tasks it is good at. The numbers are his: 36 manuscripts across 18 fields with 19 coauthors. The BootLoops tool is open on GitHub.
- 15 elliptic Feynman integrals computed for the first time and 15 known results reproduced; 5.7 billion mutation pairs from the 1000 Genomes project.
- 4,452 economics papers with code ported from paid to open software; a word-stress database for 6,072 languages; on Barro Colorado island species turn over 4.5 times faster than neutral theory predicts.
- The author says it himself: Claude and GPT are good at science, but they are not scientists. Outside his own field he brought in an expert before trusting the result.
Thirty-six manuscripts in three months. Said by a university, it would sound like a factory. It comes from one physicist with one tool and 19 coauthors, which is why it needs slow reading.
The important part is further down. Most first results were technically correct and scientifically dull. The model finds a connection nobody asked about, and it excites nobody. Only when someone who knows where the field hurts sits down next to it does something publishable come out.
The best microscope in the world does not make a discovery. It shows. Where to point it is decided by a person.
Some will argue whether 36 manuscripts are 36 papers. A manuscript is not a peer-reviewed publication, and the text does not claim otherwise. The number that will tell the truth is how many of them pass review by spring.
Do not read this as "scientists are redundant". Read it as an instruction: find the task that fits the tool, and keep the expert beside you. That is how good work was done before there were models.