A model that learns how the world works, physics, objects and consequences, rather than how text goes, so that it can predict what happens next if you push something. The 2018 paper trained an agent inside its own dream of a video game. By 2024 it was a lab strategy, with playable generated video worlds and a company founded on the idea. Whether language models already contain one is the argument.
The phrase is from psychology before it is from AI. Kenneth Craik proposed in 1943 that the mind carries a small-scale model of reality to try out actions. The 2018 paper borrowed the term with acknowledgement. Later press releases borrowed it without.
I remember the 2018 demo where the agent learned to drive a racing game by training inside a blurry generated version of the track, and then drove the real one. It was dreaming, in a strict technical sense, and it worked. I think about it every time someone says the models do not understand anything.
Sighted in August 2025 when a research lab showed a model producing a walkable 3D world from a text prompt in real time, with the paint you sprayed on a wall still there when you turned back. The record notes that object permanence, at which toddlers succeed, was the headline feature.
For the file: a world model is a simulator inside the head. Humans have one; you know what a dropped cup does before it lands. The question of the round is whether predicting text forces a model to build one, or whether it can fake the surface forever. Both camps point at the same generated videos.
What does it mean? Write it the way you would say it out loud.
