Part I
Why world models?
What you should noticePrediction is useful when it captures how the world changes — not when it memorizes pixels.
Animals (and robots) do not need a perfect video of tomorrow — they need a compact sense of what will change if I act. That is the job of a world model.
This exhibit walks the short path from encoders and reconstruction losses to JEPA (predict in latent space) and the architecture of LeWorldModel. Each interactive chapter has one empirical punchline.