A new Anthropic guest post describes physicist Matthew Schwartz using Claude and his BootLoops toolkit to solve 15 integrals that, he says, had never been computed. The work also took the approach into fields beyond his own, including ecology, where the post reports a result about how quickly tree species change on a well-studied Panamanian island.
Anthropic Watch analysis
What happened
Schwartz, a professor of physics, says he developed BootLoops by building tools around problems suited to current language models: tasks involving coding, mathematics and results that can be checked. In work on scattering amplitudes, the toolkit produced 30 integrals end to end: 15 reproductions of known results and 15 he says were new. He says this took a few weeks.
The post also describes applying related methods beyond theoretical physics. In ecology, Schwartz says BootLoops solved an equation used to test neutral biodiversity theory, and that analysis found tree-species composition at Barro Colorado Island changes 4.5 times faster than the theory allows. Read Schwartz’s account.
Why it matters
The interesting proposition is not that Claude has become a scientist in the human sense. Schwartz’s account instead describes a division of labour: the model helps adapt code and mathematical methods across fields, while domain experts judge whether a technically correct result matters scientifically. That distinction is crucial when a model can make connections faster than a researcher can assess them.
If the approach holds up, reusable tools and expert steering could make difficult, checkable calculations more accessible across disciplines. But the ecological result and the reported new calculations are claims in Schwartz’s account, not a substitute for seeing the methods and expert scrutiny behind them.
Our read
This is a stronger case for AI-assisted research than a clever demo: the post gives concrete outputs, including 15 reported new calculations, and explains how the work was checked and steered. The most useful takeaway is also the least magical one: let the model range widely, then bring in people who know what a worthwhile result looks like. Science still needs its subject experts; they may simply spend less time translating code between fields.
What to watch
- Whether the 15 reported new integral calculations are published with enough detail for other researchers to reproduce them.
- How domain specialists assess the ecological result and its implications.
- Whether BootLoops and its scientific protocols are made available for researchers using other models.
Discussion spark: For AI-assisted science, which should carry more weight: a result that can be checked mathematically, or independent experts’ judgement that it matters to the field?
Sources and evidence
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.