Simulating Tradeoffs in AI Organisation
The article suggests that simulating tradeoffs between AI organizational structures like singleton versus multipolar setups, or hidden versus known agents, is currently underexplored. The author proposes that large language models could serve as a sandbox for testing these dynamics, and asks if this view is incorrect.
Background
- The poster asks whether anyone is using LLMs to simulate different ways of organizing AI systems: one single powerful AI ("singleton") vs. many competing AIs ("multipolar"), transparent vs. hidden agents, homogeneous vs. diverse agents.
- These are central questions in AI safety and governance. A singleton (term coined by philosopher Nick Bostrom) concentrates all power in one AI, avoiding conflict but risking catastrophic errors or authoritarian control. A multipolar world has many AIs competing or cooperating — more resilient but prone to arms races and instability.
- The poster's insight: LLMs now let us cheaply run multi-agent simulations, making these abstract tradeoffs empirically testable. They're asking if this is already being done or if they've missed existing work.