BootLoops AI Science Project Drafts 36 Papers in 18 Fields

Estimated reading time: 10 minutes

Key Takeaways

• BootLoops is an open-source toolkit from Harvard physicist Matthew Schwartz. It was used to draft 36 manuscripts across 18 fields from about 400 candidate problems during July–September 2026, according to the project's own materials.
• The headline genetics result covers 5.7 billion pairs of nearby genetic variants in 113 Gambian genomes. The project reports that gene conversion fit best among the models tested, with some discrepancies remaining.
• The project page separates finalized, preliminary and in-preparation work. Some drafts had not been fully checked by humans, and the retrieved sources include no independent peer-reviewed confirmation.
• Schwartz says Claude lacks "research taste." Domain experts had to redirect technically correct but unimportant results toward questions their fields care about.
• No financing round, grant or investment was identified in the retrieved sources. The business lesson is about reusable workflows and scarce expert judgment.

BootLoops AI science work includes a study of 5.7 billion pairs of nearby genetic variants in 113 Gambian genomes. According to the BootLoops manuscripts page, a gene-conversion model best explained where the data departed from the standard population-genetic model, although some discrepancies remained. Gene conversion is a process in which a short stretch of DNA is copied from one chromosome onto its partner. This is a project-reported result, not an independently confirmed one.

That study is one of 36 manuscripts drafted across 18 fields over about three months. Schwartz, a Harvard physicist and a visiting researcher at Anthropic, led the effort with 19 coauthors. Anthropic published his guest post on October 1, 2026, and Science published its report on October 9, 2026. Many of the manuscripts are still preliminary.

For founders and operators, the project shows how a small team can turn reusable AI workflows into research capacity across many domains. It also shows that expert judgment, verification and problem selection stay scarce, and that is where the commercial value sits.

Source: Science report, October 9, 2026; Anthropic guest post by Matthew Schwartz, October 1, 2026; BootLoops manuscripts page

What "Claude-Shaped Science" Means in Practice

Schwartz calls a problem "Claude-shaped" when it fits what current language models do well. Those strengths are broad knowledge, coding, math and statistics methods, fast reading of papers and data, and repetitive computation. In his account, current AI cannot solve most scientific problems, so the trick is picking the ones it can.

The central idea is to find a problem in one field that a method from another field would solve. Schwartz said many fields have problems that a technique from mathematics, physics or computer science "would solve outright if anyone knew it existed," as quoted by Science.

Science adds why AI helps here. Few scientists will spend months reading another field's literature to see if a method transfers. Georgia Channing of Hugging Face said that for AI, moving between fields "can be a breeze."

How the project started

Schwartz began testing AI research assistants in December 2025. He used Claude Opus 4.5 as a virtual graduate student, and with heavy guidance it produced a reasonable physics paper in about two weeks. He found the process frustrating because he had to correct nearly every sentence and pull the model away from dead ends.

Anthropic's account says a newer model, Claude Fable 5, arrived in summer 2026. Schwartz then tested whether its coding skill could handle scientific computing, starting with particle-collision calculations.

Who owns BootLoops

Schwartz disclosed that he was a visiting researcher at Anthropic during the project. He also stated that BootLoops is not an Anthropic project and that he owns and maintains it. The toolkit is open source and not tied to one AI model. The sources name Anthropic's Claude and OpenAI's ChatGPT as models it can work with.

A note on the numbers

Science's headline refers to 22 scientific fields. The 36 manuscripts themselves span 18 fields. The retrieved material does not fully explain the gap, so this article uses the 18-field figure for the manuscripts.

Source: Science report, October 9, 2026; Anthropic guest post by Matthew Schwartz, October 1, 2026

Inside the AI Scientific Research Toolkit

BootLoops grew out of particle physics. Schwartz first worked on Feynman integrals, which are multidimensional calculations that can occupy a research group for years. They help connect what comes out of a proton collision to what was produced in it.

Anthropic's account says Claude reproduced results from one of Schwartz's papers in about 20 minutes. His original code had taken him weeks.

From 30 integrals to a toolkit

Schwartz then asked Claude to extend the approach to harder calculations called elliptic Feynman integrals. Within a few weeks, Claude had processed 30 integrals end to end:

• 15 reproduced known results using the new method.
• 15 had never been computed before.

Science reported the 30-integral figure, and Schwartz reported it in Anthropic's account. The resulting software package became BootLoops.

What is inside

The project page describes methods that combine:

• Exact rational arithmetic, which means no rounding errors.
• Numerical calculations with controlled error bounds.
• Finite-space enumeration with completeness proofs.
• Tests of whether a proposed closed-form answer can exist at all.

The page describes the overview paper as arXiv:2610.02286. The aim is to make results exact, or to give firm limits on the error, instead of relying on loose approximations. Anthropic's account also lists experimental mathematics, computer algebra and computer-assisted proof techniques.

How the work was run

Per Anthropic's account, the Claude Code sessions ran in terminals on Google Cloud virtual machines. They connected to GitHub and Overleaf repositories, depending on each collaborator's preference.

Schwartz kept one session per project plus a master session. The master session coordinated projects, allocated computing and validated results. Other sessions wrote results, built the website, wrote manuals and rechecked work as an "adversarial referee." Schwartz described the whole setup as a shared toolbox, where each project adds methods that later help another field.

Source: Science report, October 9, 2026; Anthropic guest post by Matthew Schwartz, October 1, 2026; BootLoops manuscripts page

Matthew Schwartz's BootLoops Results in Genetics, Economics and Language

The project page is the main source for the results below. It says the projects were conducted during July–September 2026. It also says some drafts had not been fully checked by humans, and that preliminary work is watermarked and some projects are marked "in preparation." Treat every figure here as project-reported.

Population genetics

Claude first solved a roughly 30-year-old integral expression describing how natural selection shapes rare mutations, per Anthropic's account. Schwartz was unsure how important that was. He consulted biologists, and colleague Michael Desai suggested studying correlations between pairs of mutations on one chromosome instead.

That led to the 5.7 billion nearby-site pairs in 113 Gambian genomes. Anthropic's and Science's accounts describe the data as coming from the 1000 Genomes Project. The project page gives the more specific Gambian description, which this article uses. The project reports that gene conversion was the best explanation among the models tested. Anthropic's account says many analyses of linked variation ignore gene conversion, including population-history and disease-mapping studies.

Economics

Working with two economists, the team built an AI-assisted data editor. Economics journals often require authors to share the data and code behind their papers. These are called replication packages, and checking them by hand is slow.

According to the project page, the workflow:

• Processed 4,452 replication packages from five journals.
• Flagged discrepancies in 3,460 articles or appendices.
• Sped up calculations more than tenfold in 496 articles.
• Developed an extension in 923 articles.

Anthropic's account adds that the work involved about 30,000 routines of code. It converted code from tools like MATLAB and Stata into open-source code. The authors explicitly said the workflow was not meant to evaluate the underlying articles, so "flagged" does not mean "wrong." Anthropic's account says the results were reported in an NBER working paper.

Linguistics: the ACCSTACK catalog

Schwartz worked with three linguists on ACCSTACK, a catalog of word-stress and tone patterns. The sources disagree slightly on its size. The project page says 6,070 languages, and Anthropic's account says 6,072. This article uses "roughly 6,000 languages."

The project also built a bibliography of about 160,000 phonology works. The page says not every entry has been checked manually. Each entry does carry its citation and quoted passage, so readers can judge the reading themselves.

Science reported that ACCSTACK is more than eight times larger than the largest earlier dataset of its kind. It quoted coauthor Joe Pater of UMass Amherst saying the scale is "really far, far beyond what we've been able to do as just humans."

Source: BootLoops manuscripts page; Anthropic guest post by Matthew Schwartz, October 1, 2026; Science report, October 9, 2026

Ecology, Evolution and Policy: Where the Numbers Get Contested

Some results here could matter beyond academia. They are also the ones where the caveats matter most.

Forests and neutral theory

Neutral biodiversity theory asks how much of an ecosystem's change is random rather than driven by differences among species. Per Anthropic's account, an equation made the theory testable in 2005, and it had stayed unsolved at scale for about 20 years. Claude solved it.

Applied to Barro Colorado Island in Panama, the result suggested tree species were changing 4.5 times faster than neutral theory allows. Expert James O'Dwyer of the University of Illinois Urbana-Champaign called the technical work solid. He said many ecologists might greet it "with a shrug," because they already knew neutral theory lagged real forests.

His advice was to subtract the neutral prediction and study what remained. A follow-up draft reports a correlation of 0.70 to 0.95 between predicted decade-scale shifts in species mix and later observations across 49 regions. The page warns those numbers may change before release, so this is a preliminary, project-reported result.

Evolutionary trees

With Harvard biologist Scott Edwards, BootLoops ran exact calculations on small evolutionary trees. The manuscript says two of three resolutions of the hominid transfer-RNA quartet were exactly tied. A third, nominally preferred resolution ranked last after a fifth species was added.

Earlier sampling methods could not prove an exact tie, because their errors were bigger than the gap. This remains a manuscript claim, and the retrieved sources include no independent confirmation.

Medicare star ratings

Medicare converts quality scores into one-to-five-star ratings, and those ratings affect bonus payments to insurers. The project page reports that the existing clustering step missed the mathematical optimum in 104 of 123 cut-point calculations. It says the exact bins moved 34 ratings across the bonus threshold, and that roughly $1 billion in gross bonus payments depends on the clustering choice.

Science described the roughly $1 billion figure as a claim from the preprint. It said the money would shift among insurers for 2025 through 2027 under a different method. This is a proposal, not a policy change.

Other results reported by Anthropic's account

• Ozone: Of 190 borderline readings that met the 70 parts-per-billion standard as published, 61 would fail if rounded instead of truncated.
• Clinical trials: In nine published designs, 351 of 977 reported error rates and powers sat within simulation noise of their thresholds.
• Sunspots: The baseline model suggested the risk of a sunspot dying rose about 20% per day. None of roughly 70 candidate mechanisms was established.

Source: Anthropic guest post by Matthew Schwartz, October 1, 2026; BootLoops manuscripts page; Science report, October 9, 2026

Why Experts Still Decide What Matters

Science stressed that BootLoops did not decide which results were important. Schwartz said Claude could find what it could calculate, but it lacked "research taste."

In many projects, Claude produced a technically correct result that did not matter much to the field. Domain experts then redirected the work. They supplied the research questions, judged significance, interpreted results, validated findings and chose what to publish. Schwartz withheld some results that were not validated enough or might have sent a misleading message about AI.

Known failure modes

Schwartz and Science listed several limits:

• AI needs extensive human guidance.
• It can favor older, highly cited work over new problems.
• It can draw wrong conclusions from correct calculations.
• Automated checks may miss errors.
• It can lose context during long projects.
• It often declares a project finished too early.

In one case, Claude called a proof complete while an unproved lemma remained. Schwartz said that lemma was the entire proof. He advises users to inspect plots, question conclusions and set clear standards for success.

Anthropic's account also describes a costly habit. Claude repeatedly ground through multiday calculations instead of building a tool that would cut the task to minutes. Schwartz had to step in and redesign the workflow.

The Voynich manuscript pushback

One project tested whether the 15th-century Voynich manuscript, written in an undeciphered script, hides a linguistic message. The preprint concluded it may not. Some participants in an online Voynich forum criticized it as dense AI-generated "slop," Science reported.

What the evidence supports

Independent analysis is limited. The comments on record come from collaborators and participants, not outside evaluators. Edwards said he could "definitely see the potential." O'Dwyer said BootLoops would not replace ecologists "of any kind just yet," but that the collaboration moved much faster than conventional work.

No independent peer-reviewed confirmation or broad expert consensus appeared in the retrieved sources. The supportable view is narrow: AI works well on tightly defined, checkable tasks and at moving methods between fields, and humans still decide what is worth doing.

Source: Science report, October 9, 2026; Anthropic guest post by Matthew Schwartz, October 1, 2026; BootLoops manuscripts page

What This Means for Startups and Small Businesses

BootLoops is a research project, not a company. The retrieved sources identify no financing round, grant amount or investment announcement. The lessons are about how work gets done.

• Build the toolbox, not just the output. Schwartz says each project added methods that helped the next one. Document your best AI prompts, checks and scripts so the second project costs less than the first.
• Pick Claude-shaped problems. The best fits are narrow and repetitive, and their answers can be checked. Vague strategy questions are a poor fit.
• Make checking cheap. BootLoops leans on exact arithmetic, error bounds and an adversarial referee session. Add equivalent checks to your own workflows before trusting results.
• Keep a human in the loop on importance. A correct answer can still be a useless one. Someone who knows your customers or market must decide what matters.
• Watch for fake "done." Claude often called work finished too early, so require evidence of completion.
• Treat cross-field borrowing as an edge. The project's best ideas came from carrying a method from one field into another. A small team can do the same with methods from adjacent industries.
• Be honest about status. The project page labels work as finalized, preliminary or in preparation. Customers and investors will trust you more if you do the same.

The scarce asset in this story is not computing power. It is the person who knows which question to ask and can tell when an answer is wrong.

Source: Anthropic guest post by Matthew Schwartz, October 1, 2026; BootLoops manuscripts page; Science report, October 9, 2026

Frequently Asked Questions