Anthropic blog: Harvard physicist builds BootLoops by finding 'Claude-shaped' science problems
burny_tech · x · 2026-10-02
Anthropic's science blog features a guest post by Harvard physics professor Matthew Schwartz on doing 'Claude-shaped' science: instead of forcing LLMs into traditional workflows, seek problems that match current model capabilities.
Key points:
- He argues the frustration researchers feel stems from an impedance mismatch: models excel at fast, exactly-checkable quantitative computation, while most science isn't like math where problems can be fully stated and answers absolutely verified.
- This led him to build BootLoops, a toolkit for exact calculations in quantitative science.
- Similar calculations recur across ecology, population genetics, and a dozen other fields; Claude surfaced cross-domain connections that were technically correct but scientifically unremarkable at first, so he worked with domain experts to steer BootLoops toward questions those fields actually care about.
- It follows his earlier 'Vibe Physics' piece comparing Claude's capability to a physics grad student.
More from AGI Musings
- François Fleuret: AI will force us to decide who can understand what — francoisfleuret · 2026-10-02
- Economist's EA cover story reignites debate: was EA too soft on AI risk? — AaronBergman18 · 2026-10-02
- Musk: 'Super Intelligence' Now Aces Accounting Tests AI Failed 18 Months Ago — elonmusk · 2026-10-02
- Terence Tao: Today's AI doesn't feel like intelligence, it's more like a trick — rohanpaul_ai · 2026-10-02
- Researcher JJitsev: Frontier Labs Fear Collapse of Their Trillion-Dollar Exclusivity Claim, Not Extinction — JJitsev · 2026-10-02
- Open weights will be good enough: why frontier AI smarts won't matter for daily life — Dan_Jeffries1 · 2026-10-02