Jargon:World model
1 min read
Say this at the party
“An AI's inner picture of how the world works, so it can imagine what happens next before it acts. Like playing the move out in your head first.”
In plain English
You know that if you drop a glass, it falls and probably breaks. You don't need to try it. That's your own world model.
In AI, a world model is a system that learns how things change and move, so it can predict what will happen next. Some can even generate a whole explorable world from a prompt.
Many researchers think this is a missing piece for robots and AI that act in the real world, not just chat.
Where it came from
The idea of an AI learning a model of its surroundings is decades old. A 2018 paper simply titled "World Models" by David Ha and Jürgen Schmidhuber popularised the term, showing an AI that could train inside its own "dream" of a video game. In August 2025, Google DeepMind announced Genie 3, which makes worlds you can walk around in real time.
See it in action
A robot in a warehouse needs to pick up a wobbly stack of boxes. Using a world model, it can try a few ways in its head and pick the one least likely to make the stack topple.
What people get wrong
- It isn't a map or a database of facts. It's a learned sense of cause and effect, and it can be wrong.
- A video that looks realistic doesn't prove a model understands physics. Generated worlds still drift and break the rules after a while.
Go deeper
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.