Against Possible Worlds
A viral LessWrong essay argues the 'possible worlds' framework for probability is a misleading intuition pump for AI.
A viral essay titled 'Against Possible Worlds' on the rationality forum LessWrong has sparked discussion by critiquing a foundational metaphor in probability theory and its philosophical application to AI. Written by user Ape in the coat, the post argues that the 'possible worlds' framework—where probability is conceptualized as degrees of belief about which of many logically consistent universes is actual—is a flawed 'intuition pump.' While useful as a story, the author contends philosophers and some AI theorists mistakenly treat this leaky metaphor as literal metaphysics, building complex, untenable systems on its shaky foundation.
The core critique is practical: the framework is computationally impossible for any real agent, human or AI. To reason about a simple coin toss (Ω = {Heads, Tails}), the framework allegedly requires considering every fact in every logically possible world, an infinite task. The author advocates for a saner, operational view of probability tied to real-world experiments and semantic agreements, not metaphysical speculation. This debate matters for AI development, as the models we use to formalize reasoning and uncertainty directly impact the design of machine learning systems, agents, and alignment research.
- Critiques the 'possible worlds' framework as a misleading metaphor for probability, comparing it to the outdated planetary model of the atom.
- Argues philosophers erroneously treat the intuitive story as literal metaphysics, creating confused and impractical models for reasoning.
- Highlights the computational impossibility of the framework for real-world AI, using a simple coin toss example to demonstrate its untenable infinite regress.
Why It Matters
For AI builders, the frameworks used to model reasoning and uncertainty directly impact system design, efficiency, and alignment.