An AI model that predicts how the world will change if you act in a certain way.
A world model is an AI model trained to predict "if things are like this now and I do that, what happens next?" It's like playing out a small copy of the world in your head first.
World models let robots or self-driving cars weigh up results before bumping into anything, or practise in a virtual world. In 2022 Yann LeCun proposed an AI design built around such a predictive model, and in 2025 he left Meta to start a world-model company.
In August 2025, Google DeepMind's Genie 3 turned a written description into a virtual world you could walk around in real time, staying consistent for a few minutes. Keeping track of earlier states for longer remains unsolved.
Think of it like this
Before you roll a bowling ball, you picture it: "at this angle, the right-hand pin will be left standing." A world model lets an AI do that kind of picturing ahead of time.
For example
A research demo where you type "rough path through a volcanic area" and get a virtual world you can walk along yourself is an example of a world model.
Common mix-ups
Mix-upA world model is just an image generator that makes convincing scenes.
ActuallyThe point is predicting what changes when you do something. In 2024 Genie learned only from moving scenes on the internet, with no labels saying who pressed which button, and still made virtual worlds people could control one frame at a time.
Mix-upWorld models are just toys for making games.
ActuallyIn January 2025 NVIDIA released Cosmos, world models to help build robots and self-driving cars. NVIDIA said they were trained on 20 million hours of real-world footage of people, industry, robots and driving, and they're used to make virtual practice material for training robots.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- If an LLM predicts the next word, what does a world model predict?
- Activity
- Before throwing a ball, sketch on paper where you think it will go, then throw it and compare. Next, read LeCun's 2022 proposal (E0159), Genie (E0333) and Genie 3 (E0658), and discuss what else would be needed for predictions to hold for more than a few minutes.
Records about this
Source-checked moments from AI history.
Generates controllable environments from a prompt, staying consistent for several minutes at 720p (720 lines) and 24 fps
Productized synthetic-data and simulation infrastructure for training physical AI
The starting point of the world-model lineage. The authors argue that at this scale it can be regarded as a foundation world model
The founding document of the JEPA and world-model approach. LLM skepticism took the form of an alternative design rather than criticism