What's this about?
Physicist Matthew Schwartz built a tool harness for exact quantitative work, then used Claude's breadth to spot mathematical methods that transfer across scientific fields. Domain experts remained essential for selecting questions that matter.
What exactly is a “Claude-shaped” problem?
It is a task aligned with current model strengths—searching broad literature, coding and checking formal calculations—rather than deep conceptual judgement that requires tacit scientific taste.
Why does this matter?
This is a more grounded path to AI-assisted science than asking a chatbot to “make a discovery.” It also reports the limitation honestly: technically correct cross-field connections can still be scientifically uninteresting.