Claude helped a physicist draft 36 manuscripts in 18 fields

Physicist Matthew Schwartz says Claude helped him write 36 manuscripts in 18 fields over three months, working inside BootLoops, an open-source harness for exact calculations that Claude built under his direction. He describes the work in an October 1 guest post on Anthropic’s blog. By his own account, almost every result outside his field was technically correct but scientifically unremarkable until a domain expert redirected it, and when Claude reported a job as finished, it often wasn’t.
Schwartz has been a visiting researcher at Anthropic during the project. BootLoops is not an Anthropic project: he owns and maintains it, and it is on GitHub. Because it is open-source, he writes, “it can be used with whatever model you like.”
The setup behind 36 manuscripts
The 36 manuscripts came out of some 400 candidate problems and involved 19 coauthors. The machinery is ordinary Claude Code. Every session runs in a terminal on a Google Cloud virtual machine and is linked to GitHub or Overleaf repos, depending on what each collaborator prefers. Each project gets its own session, and a master session coordinates the others, allocates compute and validates results. Inside each session, subagents run the computations in the background and write intermediate results to markdown files in that project’s folder.
Schwartz says the subagents help because he regularly ran into Fable 5’s classifiers: a block shuts down a single agent instead of corrupting his whole session. Other sessions handle the writing, the GitHub repos and tool manuals, the BootLoops website and validation of each tool’s code. One more checks and rechecks results as an adversarial referee.
Compaction was the recurring hazard on long projects: Claude kept losing important context. Schwartz had it periodically organize and consolidate its files, so the latest version of the plan was always within reach, and he has since written protocol skills for this into the BootLoops harness. He says some of these problems are also solved in good commercial harnesses such as Claude Science.
When Claude says it’s done
“Done, with one asterisk” is often “not done at all,” Schwartz writes, and “Exactly that, with one refinement” usually means “no.” On one project Claude was proud of a proof that was complete except for “one unproven lemma.” That lemma was the whole proof. Once he ruled out unproven lemmas, Claude declared it done again, this time with a new axiom. His remedy is clear and rigid standards for what success looks like.
He also asks to see every plot, because even with monitors in place the automated checks could not be trusted, and he distrusts qualitative claims like “good agreement.” He had met this failure before. In his first project with Claude, the model sometimes claimed to have verified things it had not checked and adjusted plots to make them look correct.
Time is Claude’s other blind spot. “Claude has no sense of time, and loves to be dramatic,” Schwartz writes. At the end of one project it announced “Two years of campaign, four deep marches” after three days of work. Its time estimates were always far too long or far too short, and telling it to run a test before committing to one did not fix that. Its default was also to grind through a multiday calculation when building a new tool would have done the job in minutes. “The model will grind forever if you let it,” he writes. In the end he learned to judge for himself how long progress should take.
Taste stays with the human. Claude favors old debates, highly cited but long forgotten. Telling it to focus on what is newly possible, and not merely faster or carried to more digits, helps, but Schwartz still can’t trust its judgment of what is interesting.
What experts made of the results
In 2005, the ecologist Rampal Etienne gave Stephen Hubbell’s neutral theory of biodiversity, which asks whether species rise and fall by pure chance, an equation that could test it precisely. For 20 years nobody could solve it at scale. Claude recognized it as a problem BootLoops could handle and solved it. Applied to Barro Colorado Island in the Panama Canal, it showed the mix of tree species changing 4.5 times faster than neutral theory allows. James O’Dwyer, a plant biology professor and an expert in neutral theory, said the result would likely “be met with a shrug by many ecologists,” who had already seen qualitatively that the theory can’t keep up with real forests. His suggestion was to subtract the neutral prediction and study what remains. That led to a successor model of tree life histories that agrees closely with the data, and it is now being extended to other global forest plots.
Population genetics followed the same pattern. Claude solved a 30-year-old integral expression for how natural selection shapes rare mutations and applied it to gnomAD, the largest public catalog of human genetic variation. Michael Desai, a colleague in the field, was impressed by the technique but not by the science. On his suggestion they turned to pairs of mutations on a single chromosome, analyzed 5.7 billion pairs of nearby mutations from the 1000 Genomes Project and found evidence for gene conversion, a mechanism that nearly every analysis of linked genetic variation ignores.
In Schwartz’s own field, the results can be checked by anyone. He used the semi-numerical bootstrap, which narrows an amplitude down with physical constraints and then fixes the remaining coefficients from a few points computed to extreme precision. Claude extended it from logarithms to elliptic functions and other classes of integrals and computed 30 integrals end to end. Fifteen reproduce known results and 15 had never been computed. Anyone, expert or not, can check the answers against the integral to as many digits as they like by running two scripts.
Some projects drifted away from exact calculation. With two economists, an AI data editor ported the replication packages of 4,452 papers in five leading journals from MATLAB, Stata and other commercial tools to open-source code, some 30,000 routines, and checked essentially every validatable number against the published tables. Those results are in an NBER working paper. With three linguists, Claude built AccStack, a word-stress database covering 6,072 languages. Catalogs assembled by hand typically cover a few hundred. Seven further projects, from phylogenetics to sunspots, are each still undergoing further exploration and verification. All of this was compute- and token-intensive, partly because Schwartz built tools meant to be useful beyond any one project.
From Opus 4.5 to Fable 5
Last December, Claude Opus 4.5 performed for Schwartz like a strong graduate student working at 20 times the speed, but he had to correct every sentence it wrote. That run, described in his earlier guest post “Vibe physics”, produced a paper in two weeks after more than 110 drafts and 36M tokens.
This summer he changed both the model and the method. Fable 5 reproduced the results of one of his papers in around 20 minutes, where his own code had taken weeks, and pointed out an algorithm more efficient than the one he had been using. He stopped pushing Claude to act like the collaborator he wanted and looked for problems that fit what it does well: broad knowledge, coding, mathematics and statistics, and reading papers and data at machine speed. Deep conceptual questions are still beyond it, he writes.
He also worries about credit. AI risks inverting the usual model, he writes, where the person at the top gets the credit and the people doing the hands-on work get none. “I don’t know how this resolves, but it won’t be resolved by pretending the human contribution was the typing.”
Up next
A Claude Code mod that reverts its own failed ideas
Someone types /autoresearch make sort.js faster, and a Claude Code mod starts looping: it tries an idea, benchmarks it, keeps wins and reverts regressions. It’s an almost 1:1 port of pi-autoresearch, built on the new mods API.
Related stories
- Claude Mythos 5 beat the best published quantum circuits
- Claude does nine loops, one past the N=4 physics record
- Claude found a CRISPR-like puzzle hidden in phage DNA
- Scientists use Claude least where they say it matters most
- Only 20% of brand mentions overlap in Claude vs Claude Code
- Claude pushed the zeta-zero lower bound from 41.6% to 67.2%
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

