Skip to content- Sources: primary, discussion, discussion
- Summary: Matthew Schwartz reports selecting problems that suit what current models do well, building an open-source toolkit called BootLoops, then recruiting domain experts to steer each result toward something the field values, and he states the first findings were technically correct but scientifically unremarkable until an expert redirected them. Reported scope is 36 manuscripts across 18 fields with 19 coauthors over three months, from roughly 400 candidate problems. This is a guest post hosted by the model vendor, the findings are author-reported rather than peer reviewed, and several are stated as still undergoing verification.
- Why it matters: The described method inverts the usual framing by fitting problem selection to current model strengths instead of asking the model to work like a scientist.
- Follow-up: Whether any of the 36 manuscripts clears peer review.