LeMario: Training a JEPA World Model on Super Mario Bros

I built LeMario from scratch to test if a JEPA model could learn Super Mario Bros dynamics without rewards. While the model predicted short-term futures accurately, it failed to navigate complex obstacles. My experiments revealed that predicting the game is not the same as learning how to make progress through it, offering a crucial lesson on the limits of current world models.

"The model had learned to predict the game, but that did not mean it had learned how to make progress through it."