@dmarzzz
@cytonomy
@Sasank-C
--:--
0 runs in flight
the highlights
distributional safety research program
Most AI safety work tests one model at a time. Once agents trade, talk, share memory and hand off work, capability and risk live in the group, and a swarm can fail in ways none of its members would alone. Distributional safety asks which safety properties hold for the swarm as a whole. We tested seven of them, in controlled experiments and in archives left by real agent swarms.
The term comes from Google DeepMind's Distributional AGI Safety (2025), which argues general capability may first appear as a patchwork of many cooperating agents.
Click a track on the dial to jump to its results.