World models explained simply: the next big idea in AI

Chatbots learn to predict the next word. A world model learns to predict what happens next in a world. In late 2026, world models are one of the hottest ideas in AI research. Google DeepMind released Genie 4 in October, and Yann LeCun, one of the pioneers of deep learning, started a new company to build them.

Below: what world models are, and how to start learning about them.

What is a world model?

A world model is an AI model that learns how an environment works. Give it the current state of the world and an action, and it predicts the next state.

Example: you are playing a car racing game and you press "turn left". A world model predicts the next picture on the screen: the car turns and the road moves. It learned this by watching many examples, not from rules of physics written by hand.

People do this all the time. Before you catch a ball, your brain predicts where it will be. Researchers want AI to have the same kind of inner model of the world.

World model vs language model

• A language model predicts the next word in a text.
• A world model predicts the next state of an environment, often as video, and it takes actions into account.
• A language model learns what people wrote about the world. A world model tries to learn how the world behaves.

Where the idea comes from

The idea is old, but a 2018 paper made it popular in deep learning. In "World Models", David Ha and Jürgen Schmidhuber trained a model on a simple car racing game. The model learned a small summary of the game and how it changes over time.

Then they did something surprising: they trained a driving agent entirely inside the model's own "dream", and it still worked when they moved it back to the real game. The paper has an interactive version at worldmodels.github.io. It is a great first read.

Two ways to build a world model

• Generate the world: the model makes the next video frames you would see. Google DeepMind's Genie models work this way.
• Predict the meaning: the model predicts a short summary of what comes next, not every pixel. This is the "JEPA" idea that Yann LeCun has pushed for years. Meta's V-JEPA 2 (2025) learned from over 1 million hours of internet video, and was then used to plan robot actions.

Genie 4: what Google released

On 27 October 2026, Google DeepMind announced Genie 4, a world model for real-time, interactive simulation. You can steer a scene with plain language, for example move the camera, change the weather or add a character. It also makes sound for each world as you go, including music and dialogue.

Genie 4 is an early research preview, and only trusted testers can use it now. Google says it plans wider access later. DeepMind also says it cannot yet simulate real places with perfect accuracy.

Why Yann LeCun is betting on world models

Yann LeCun won the Turing Award, often called the Nobel Prize of computing, for his work on deep learning. He left Meta at the end of 2025 and started AMI Labs (Advanced Machine Intelligence), based in Paris, to build world models. In March 2026, AMI Labs raised about $1.03 billion.

LeCun has long argued that language models alone will not reach human-level intelligence, because they do not learn how the physical world works. AMI Labs has not released a public model yet.

Why world models matter

• Robots: a robot can practise in a simulated world before it acts in the real one. That is cheaper and safer.
• Self-driving cars: rare and dangerous situations can be tested without real risk.
• AI agents: an agent can "imagine" the result of an action before it acts.
• Games and learning: interactive worlds made from a simple text prompt.

The research is just starting

Many problems are still open: keeping a world stable over a long time, getting the physics right, and measuring how good a world model really is. NeurIPS 2026, one of the biggest AI conferences, has a full workshop called "World Models in Physical AI" in December. This field needs new researchers.

How to start as a student

1. 1Read the interactive "World Models" paper at worldmodels.github.io.
2. 2Learn the basics of reinforcement learning: states, actions and rewards.
3. 3Try a small project: train a model to predict the next frame of a simple game from the free Gymnasium library.
4. 4Then read the V-JEPA 2 paper to see the other approach.

At APJ-RAISE, students read papers like these together and turn them into small research projects. Joining is free.