Matt Schwartz (Harvard): Bootloops Claude-shaped Science Across Disciplines

Abstract: BootLoops is an open-source harness for large language models aimed at exact, checkable quantitative science. Originating in scattering amplitudes and Feynman integrals, it seeks problems whose mathematical structure recurs across disciplines, from cosmology and string theory to population genetics, ecology, genomics, and linguistics. I will describe the principles behind BootLoops – checkability, breadth, tool-building, and adversarial vetting – and discuss how AI can help connect distant regions of scientific knowledge by finding ‘Claude-shaped’ problems that are especially well-suited to current models.