A tiny, fully-reproducible JEPA world model that learns the physics of a bouncing DVD logo in representation space, dreams its future, and detects anomalies. Trains on a CPU in ~10s. Interactive browser demo.
machine-learning deep-learning reproducible-research pytorch representation-learning anomaly-detection self-supervised-learning video-prediction world-models interactive-demo jepa i-jepa v-jepa
-
Updated
Jun 13, 2026 - Python