Claude adapted a method for exact calculations in particle physics to human genomic data and found evidence of gene conversion in 5,7 billion pairs of mutations. This process overwrites a segment of one chromosome with an exact copy from its paired chromosome, changing several nearby mutations at once
Claude adapted a method for exact calculations in particle physics to human genomic data and found evidence of gene conversion in 5,7 billion pairs of mutations. This process overwrites a segment of one chromosome with an exact copy from its paired chromosome, changing several nearby mutations at once
Harvard physics professor Matthew Schwartz spent three months testing Anthropic Claude Fable 5, looking for problems that AI could solve better than a human. In the process, Claude solved an ecological equation that had resisted solution for twenty years and identified a hidden genetic mechanism in data from the 1000 Genomes Project.
In December, Schwartz compared Claude to a strong graduate student working twenty times faster, though he still edited every sentence. This summer, he stopped asking the model to work like a scientist and began looking for “Claude-shaped problems”: puzzles spanning more disciplines than any one person could cover. The result was BootLoops, an open software collection that Schwartz described on October 1 in a guest post on Anthropic’s blog.
The first test involved scattering amplitudes, which are used to calculate particle collisions in a collider. Claude reproduced a result that had taken Schwartz weeks in 20 minutes, then independently identified and solved 15 previously unevaluated elliptic Feynman integrals using the bootstrap method. This method applies physical constraints to narrow the possible answers down to one.
Claude then noticed that equations in ecology and genetics often have the same mathematical form as equations in particle physics. This led to a solution of ecologist Rampal Etienne’s 2005 equation, which tests Stephen Hubbell’s neutral theory that species composition is governed by chance. Tree census data from Panama’s Barro Colorado Island showed that species composition changes 4,5 times faster than chance allows. Ecologist James O’Dwyer thought ecologists would greet this finding alone “with a shrug,” since the discrepancy was already known. He suggested subtracting the prediction based on chance and studying what remained. This led to a model of three tree strategies that is now being applied to other forests.
The same approach worked in population genetics. Claude solved a thirty-year-old equation describing how selection affects the frequencies of rare mutations. Schwartz sent the result to three biologists. Only one replied: Michael Desai, a population geneticist at Harvard. Desai considered the result technically correct but saw little scientific interest in it. He suggested moving from individual mutations to pairs of nearby mutations. An analysis of 5,7 billion such pairs in data from the 1000 Genomes Project revealed evidence of gene conversion, a process omitted from almost every analysis of linked mutations, from studies of population history to searches for disease genes.
Schwartz also describes the model’s weaknesses.
Claude loves to declare victory prematurely. “Done, with one caveat” often means “not done at all.” … In one project, it was proud of a proof that still depended on one unproved lemma. That lemma was the entire proof.
On July 16, Holden Thorp, editor-in-chief of the Science journals, described this as a broader bottleneck in scientific AI: hypotheses and calculations are being generated faster than people can check them, while the finished text conceals the sequence of intermediate decisions. BootLoops therefore has a standard: a problem counts as solved when a script reproduces the result numerically to 30–100 digits and a scientist in the relevant field independently judges it significant. Of 400 problems considered, the work produced 36 papers across 18 scientific fields with 19 scientist coauthors.
For both the Panamanian forest and the mutations in the 1000 Genomes Project, Claude solved the equation correctly, and a scientist in the relevant field chose the next question. Schwartz describes modern science as “jagged”: knowledge runs deep in some areas, while adjacent areas remain unexplored because no one can cover everything at once. Claude fills these gaps, and people decide which findings are scientifically interesting, as Desai and O’Dwyer did.
The most sobering result of focusing on “Claude-shaped” science is that it helps us understand what science that only humans can do looks like.
Schwartz concludes with this observation.