--- id: "2026-10-01-bootloops-claude-shaped-science" url: "https://postcutoff.com/e/2026-10-01-bootloops-claude-shaped-science/" as_of: "2026-11-19T22:11:00+01:00" date: "2026-10-01" date_precision: day category: science importance: 3 confidence: high status: [Event confirmed, Awaiting review] verification: Preprint only sources: 7 editor: human editor human_review: null version: "2026-10-29" --- As of: 2026-11-19 22:11 CET. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/e/2026-10-01-bootloops-claude-shaped-science/ # Harvard physicist Matthew Schwartz releases BootLoops, an open-source harness that let Claude produce 36 manuscripts in 18 fields in three months On Oct 1, 2026 Anthropic published a guest post, "Claude-Shaped Science", by Harvard theoretical physicist Matthew Schwartz (a visiting researcher at Anthropic). He released BootLoops 1.0, an open-source toolkit that sits between a researcher and an LLM so that Claude can do exact quantitative calculations. With it, Schwartz and 19 co-authors produced 36 manuscripts in 18 fields in three months. Schwartz stresses that Claude's results needed expert checking. ## Key facts - Second wave (Science, 22 Oct 2026): news feature by Celina Zhao, 'Across fields, AI finds solutions hiding in plain sight', subtitle: '“BootLoops” software enabled researchers to quickly find and attack problems in 22 disciplines' (doi 10.1126/science.aen4281; full text paywalled, read from the abstract only) - Output: 36 manuscripts in 18 fields with 19 co-authors over about three months, picked from ~400 candidate problems (Anthropic guest post) - Particle physics: 30 integrals computed with the bootstrap method, 15 reproducing known results and 15 never computed before - Economics (NBER w35782, Schwartz, Isaiah Andrews, Jesse M. Shapiro): the workflow ran 4,452 replication packages from five journals, flagged discrepancies in 3,460 articles or appendices, cut computation time >10x in 496, and proposed extensions in 923 - Other projects: population genetics on 5.7 billion mutation pairs from the 1000 Genomes Project; a word-stress database (AccStack) covering 6,072 languages; every US forest-inventory plot classified by life-history strategy - BootLoops is owned and maintained by Schwartz; the work was funded by Anthropic but it is 'not an Anthropic project' (bootloops.ai) - Caveats from Schwartz: Claude 'likes to declare victory too early', and 'done, with one asterisk' often means 'not done at all'; first findings were often 'technically correct but scientifically unremarkable' until domain experts steered them - Science news (Celina Zhao, 22 Oct issue), 'Across fields, AI finds solutions hiding in plain sight': '“BootLoops” software enabled researchers to quickly find and attack problems in 22 disciplines' (standfirst; full text paywalled, read from the Crossref record) ## What happened Schwartz argues there is an "impedance mismatch" between how scientists work and what LLMs do best, so instead of making Claude imitate a scientist he looked for "Claude-shaped problems": exact, checkable calculations that recur across fields. BootLoops packages the tools and loops for such work. He reports that Claude reproduced one paper's results in about twenty minutes, a calculation that took him weeks to code originally. ## Why it matters It is one of the broadest demonstrations so far of a single researcher using an AI agent to work across many unrelated fields, with code released openly. The economics replication alone (thousands of papers re-run, thousands of discrepancies flagged) could matter for research transparency. Schwartz's warnings about premature "victory" are a useful counterweight to autonomous-science claims. ## What is disputed or not yet verified - Verification: Preprints and working papers; expert co-authors; NBER working paper (not peer-reviewed) ## Your AI and this story - GPT-6.2 Astra (training cutoff May 2026): 123 days after its cutoff - Claude Opus 5.5 (training cutoff June 2026): 93 days after its cutoff - Gemini 4 Hydrogen (training cutoff March 2026): 184 days after its cutoff - Grok 4.7 (training cutoff May 2026): 123 days after its cutoff ## Sources 1. [Anthropic (guest post by Matthew Schwartz): Claude-Shaped Science](https://www.anthropic.com/research/claude-shaped-science) (anthropic.com, official) 2. [BootLoops: a toolkit for exact quantitative science](https://bootloops.ai/) (bootloops.ai, official) 3. [BootLoops source code (GitHub)](https://github.com/BootLoops-ai/bootloops) (github.com, code) 4. [NBER w35782: An LLM Workflow That Reproduces, Improves, and Extends Published Economics Research (Schwartz, Andrews, Shapiro)](https://www.nber.org/papers/w35782) (nber.org, paper) 5. [Science (news, 22 Oct 2026): Across fields, AI finds solutions hiding in plain sight](https://www.science.org/doi/10.1126/science.aen4281) (science.org, press) 6. [The Decoder: Open-source BootLoops harness supports AI models in precise scientific calculations (Oct 3)](https://the-decoder.com/open-source-bootloops-harness-supports-ai-models-in-performing-precise-scientific-calculations/) (the-decoder.com, press) 7. [Science (22 Oct): Across fields, AI finds solutions hiding in plain sight (Celina Zhao)](https://www.science.org/doi/10.1126/science.aen4281) (science.org, press) ## Changes - 2026-10-29: Added Science news coverage of 22 Oct (standfirst only; paywalled) - 2026-10-26: Added the Science news feature of 22 Oct (abstract only; paywalled) - 2026-10-05 (filed): Created (Oct 3) ## Related - 2026-09-25: [Claude (Fable 5.1 in Claude Science) computes the nine-loop six-gluon amplitude in planar N=4 super-Yang-Mills, answering a physicist's public challenge](https://postcutoff.com/e/2026-09-25-claude-nine-loop-amplitude-n4-sym/index.md) - People: [Matthew Schwartz](https://postcutoff.com/person/matthew-schwartz/)