Run my cross-disciplinary research program: find and solve 'the AI-shaped' problems
Harvard physicist Matthew Schwartz built BootLoops, an open-source harness that points the AI at exact, checkable calculations, and ran it for three months across roughly 400 candidate problems: the result was 36 manuscripts in 18 fields with 19 coauthors, including 15 Feynman integrals never computed before, Watson's 1939 'final problem' solved, and a forest model showing Panama's Barro Colorado species mix changes 4.5 times faster than neutral theory allows. The AI reproduced results from his own paper in about 20 minutes that had taken him weeks to code.
How it works
- Have the AI port the scattering-amplitude methods from his papers and the adjacent literature into one framework
- Ask it to hunt for unsolved problems the machinery fits (elliptic integrals first: 30 computed, 15 of them new)
- Follow the same equations into ecology, genetics, economics, linguistics and more
- Run parallel AI sessions on cloud VMs, one per project, with a master session coordinating compute and validating results
- Bring each field's expert in to steer the finding into science that field values
- Write up 36 manuscripts with 19 coauthors
Why it’s worth knowing
One physicist plus an agent produced 36 manuscripts across 18 fields in three months — including integrals nobody had ever computed.
Where it’s from
Summarized from a public post on anthropic.com. See the original ↗