Entanglement Between an AI and Its Environment

According to LessWrong, researchers have introduced a framework called 'entanglement' to describe how AI systems relate to their environments. At its core, entanglement is the amount of information an AI needs to receive to accomplish a task. The authors distinguish between two types: actual entanglement, the information a particular AI instance gets while solving a problem, and minimum entanglement, the theoretical floor—the least amount an AI absolutely requires to succeed. The practical insight is striking: you can often solve the same task with far less information than you'd expect. To get an AI to help you play chess, for instance, you don't need a humanoid robot with full sensory input; you might just tell it the opponent's moves in text form. But there's always a limit. Strip away too much information, and the task becomes unsolvable. The authors argue this matters for AI safety and evaluation, because testing whether an AI can handle real-world problems requires test environments that provide enough detail to make those problems theoretically solvable.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton