Harvard Physicist Produces 36 Manuscripts with Claude in Three Months
On October 2, Anthropic's science blog published a guest post by Harvard physicist Matthew Schwartz titled "Claude-shaped science," describing his practice of accelerating research with Claude and his own toolkit BootLoops: over roughly three months, he screened about 400 candidate questions and produced 36 paper manuscripts with 19 collaborators across 18 scientific fields, including solving a problem that had troubled ecologists for 20 years.
Confirmed
- Schwartz used the physics concept of "impedance mismatch" to describe AI and science: both systems work well individually but match poorly—LLMs are capable, but interacting with them as human collaborators does not unlock their scientific potential.
- His method has Claude find "Claude-shaped" scientific problems—those best suited to current LLM capabilities—rather than forcing the model onto arbitrary problems. JMateosGarcia noted that Schwartz had previously judged Claude comparable to a physics graduate student in Vibe Physics.
- Anthropic also introduced BootLoops, Schwartz's toolkit for precise computation in quantitative science, addressing cases where LLMs err in rigorous numerical derivation; badumtsssst summarized it as moving computational techniques across disciplines.
- Per alexverem, Schwartz is an Anthropic visiting researcher and the project is Anthropic-funded.
- The work drew wide attention, with sharers like DrSingularity framing it as "AI doing years of science in days."
Unconfirmed
- Poster 1411337 claimed the research "replicated weeks of work in 20 minutes," "completed 15 previously unsolved physics calculations," and named the model "Claude Fable 5"—the model name conflicts with official wording (Claude), and these details are not corroborated by primary sources.
Why it matters
- The case demonstrates a pragmatic path for LLM-assisted science: rather than pursuing generic human-like collaboration, designing workflows and tools around model capability boundaries can scale research output across disciplines.
- Solving a decades-old problem suggests the workflow may not only increase quantity but also crack previously stalled interdisciplinary questions.
2026-10-02 ~ 2026-10-04 · 13 related posts
Primary sources
- Anthropic's BootLoops: a toolkit for exact calculations in quantitative science — badumtsssst · 2026-10-02
- [source] Harvard physicist uses Claude to link physics calculations to ecology, genetics — AnthropicAI · 2026-10-02
- Harvard physicist's 'impedance mismatch' framing of AI and science sparks sharp reply — AryHHAry · 2026-10-02
- Claude Did 3 Months of Cross-Field Science: 36 Manuscripts, 15 Never-Computed Feynman Integrals — Dr_Singularity · 2026-10-02
- Anthropic Details BootLoops: Prof. Schwartz Lets Claude Find 'Claude-Shaped' Science Problems — Dr_Singularity · 2026-10-02
- 'AI Is Starting to Do Years of Science in Days': Claude's Cross-Field Research Report — Dr_Singularity · 2026-10-02
- Harvard physicist's 3 months with Claude: weeks of work in 20 minutes, 36 papers — 141_1337 · 2026-10-02
- Physicist claims Claude produced 36 manuscripts across 18 fields in three months — IgorCarron · 2026-10-02
- Anthropic blog: Harvard physicist builds BootLoops by finding 'Claude-shaped' science problems — burny_tech · 2026-10-02
- [source] BootLoops: using Claude to carry exact-calculation techniques across scientific fields — JMateosGarcia · 2026-10-02
- [source] Harvard physicist's Claude experiment: 36 manuscripts, 18 fields, in three months — alex_verem · 2026-10-04
2 near-duplicate retellings: JMateosGarcia · alex_verem