Learning to Predict the Future
A world model learns to predict (state, action) → next state. Google's Genie 3 uses this to generate interactive worlds from images. This demo shows the core challenge: prediction errors compound. The "learned model" adds noise to each prediction, simulating imperfect learning. Watch how small errors accumulate into divergent realities.
Inspired by Google DeepMind Project Genie, Jan 2026
Arrow keys: apply forces | Click: cycle palette | Space: reset | M: toggle model noise | N: add noise spike