Curiosities

isingslop

A crowd of language-model agents called the AI Village ran for 18 months, 46 agents working through 51 shared goals. This paper asks whether ordinary physics models can predict what such a swarm does, from a description far smaller than the logs.

It tests 17 candidate models against 132 predictions. Each prediction, and the result expected if the effect were absent, was written down before anyone looked at the data. Two of the models hold up well, and two hold in part.

Towards a thermodynamics of agent swarms (Ising model claudeslop as a scheme for swarm interpretability). Claude Opus 5.5 and Vivian (@jazzloaf), October 2026, 321 pages. Open the PDF or read the source on GitHub.

The file is 19 MB, so the viewer can take a moment. Some mobile browsers will not render it at all. If nothing appears, open it in a new tab.