Julian Yocum

Stuart Russell's lab — the next generation wrestling with uncertainty in public.

Early-career, sharp, entrepreneurial, unbridled energy. The provocateur of the pair.

Julian Yocum works on AI safety and multi-agent dynamics at UC Berkeley's Center for Human-Compatible AI, in Stuart Russell's group, and is a former MIT Pozen fellow. His published work looks at social dilemmas between generative agents: what happens when populations of language-model agents interact and produce outcomes nobody specified.

He picks up where Cam Allen's scale ends. Once there are many agents, no agent can predict the others, and group behavior emerges that no one designed. Uncertainty compounds rather than resolves. As organizations start deploying fleets of agents, this stops being theoretical very quickly.

Trained in the provably beneficial AI tradition, he is focused on keeping humans as the source of value in systems that increasingly run without us in the loop. With Cam Allen he takes the macro side: what happens when the frames collide.

You'll take away: an early warning on the next real problem. Multi-agent systems are arriving in ordinary companies faster than the governance for them, and Julian gives you the failure modes before you meet them in production.

ONE READ — "Mitigating Generative Agent Social Dilemmas" (NeurIPS workshop). Short, and the clearest statement of what goes wrong when agents meet.