What is a world model?

An AI’s “world model” can be thought of as a map or description of reality (or a subset of reality) that the AI uses to navigate the world, predict the future, or make decisions.

During training, neural networks pick up patterns present in their training data and compress them into a set of interconnected rules and algorithms stored in their neural weights. The resulting collection is in some ways equivalent to a neural network's world model.

One difference between narrow AI and general AI (AGI) can be defined by the scope of their world model. An AI with a world model diverse enough, and complete enough, to allow it to accomplish a wide range of tasks would be worthy of being called weakly general intelligence. Arguably, models like GPT-4 are an example. However, the hallmark of general intelligence is the ability to solve problems in many areas not found in its training data. Or learn to do so autonomously. Whereas a narrow AI, regardless of how complete its world model is, can only act effectively on a specific subset of reality.