World Models: AI That Imagines
Predict the next state, feed it back in, and watch the imagined future drift
What your child can do at the end
Explain what a world model is and why an AI needs one to plan
The brief
Before you catch a ball you imagine where it will be. Self-driving cars, robots and game-playing AIs do the same: they learn a model of how the world moves, then run it forward in their head, step by step, to plan. That is a world model. In this studio the world is a bouncing ball or a swinging pendulum. You watch a sequence of states, fit a model that predicts the next state from the last few, and then press 'Imagine the future': the model predicts frame after frame from its own predictions, and you see the imagined ball drift away from the real one. You will discover that two lags are enough to imagine a pendulum almost perfectly, that a bounce breaks a simple imaginer, and that every imagination drifts eventually.
Skills it develops
How it is assessed
3 graded challenges inside the lab, and 8 questions at the end that check the idea rather than the clicks. Every attempt is saved, so a retry after a mistake counts as the learning it is.