Distributed vs centralized agents
AI researchers are shifting from single-minded agents to more resilient, team-like systems.
Deep Dive
A researcher argues the dominant model of AI as a single, perfectly rational 'centralized agent' is incomplete. He proposes studying 'distributed agents', where sub-components have more autonomy. This makes systems less efficient but more robust to unexpected situations. The goal is a hybrid 'coalitional agency' that balances both strengths. The ideas were shared in a talk, as a full write-up is delayed, highlighting an ongoing conceptual shift in AI safety research.
Why It Matters
This theoretical shift could lead to AI that is more adaptable and safer in the real world.