After 600 experiments, here's what I discovered about JEPA world models

A researcher conducted 600 experiments on LeWM, a JEPA world model that learns to predict future states in embedding space rather than generating pixels, and found that the model's predictions are already nearly as accurate as perfect physics for planning tasks. By developing four new planning methods without retraining the model, the researcher improved LeWM's success rate on long-horizon tasks, demonstrating that gains for world models come more from how the trained architecture is used rather than from improving the predictor itself.