Chapter 16 · AI in science and the physical world
World models
A world model predicts what happens next, so an agent can imagine outcomes before acting.
Chapter 16 · AI in science and the physical world
A world model predicts what happens next, so an agent can imagine outcomes before acting.
Future A: walks straight at the goal… into a pit.
Before moving, the agent uses its world model to play out possible futures. It drops the ones that end in a pit and follows the one that reaches the star.
It answers one question: “If I am here and I do this, what will happen?” A tiny one is a rule, like “moving right shifts you one square.” Big ones are neural networks that predict the next picture or sensor reading.
Tap squares to add or remove pits. Then change how many moves ahead the agent imagines and run it. At 0 it just walks towards the star.
Next move: → · best of every 3-move future
A simplified teaching toy: here the world model is a perfect rule. Real models are learned, so their predictions are only good guesses.
What does a world model let an agent do?
In the grid, what went wrong for the agent with 0 moves of imagination?
Why do agents re-plan after each real step?