What's this about?
Seth Lazar's one-line provocation, and Ethan Mollick's reply that improves it: "A political philosopher would say that. As an economic sociologist who studies organizational design, clearly multi-agent alignment is an organizational design problem with strong components of economic sociology" — adding that many academic disciplines have a lot to add here.
What exactly is multi-agent alignment?
Single-agent alignment asks how to make one AI system do what its principal wants. Multi-agent alignment asks what happens when many systems with different principals and different objectives operate in a shared environment. That's a different problem, and an old one under other names: how do you design rules so self-interested parties produce good collective outcomes, how do you prevent coordination failures, how do you allocate authority and resolve conflicts. Political philosophy, mechanism design, organizational theory and economic sociology have each spent a long time on versions of it.
Why is this worth noticing?
- The joke has a real claim inside it. Multi-agent systems are being designed by people reinventing institutional design from scratch, with no reference to the literatures that studied it.
- Mollick's addition is the better version, because organizational design is closer to the engineering reality: agents in a firm are more like employees under incentives than citizens under a constitution.
- The week supplies the evidence. Physics of Agents modelled collective agent behaviour with statistical mechanics; the Authority Resolution Framework built a governance ontology. Both are outside-discipline imports.