Yann LeCun has spent years arguing that generative AI is the wrong path for building intelligent systems. Now he's got more ammunition. A new paper from LeCun and colleagues at Meta introduces LeWorldModel, a world model that learns from raw pixels with just two loss terms and plans 48 times faster than foundation-model-based alternatives. The approach builds on Joint Embedding Predictive Architecture (JEPA), LeCun's preferred framework for world modeling. Instead of generating pixels, you predict abstract latent representations. Previous JEPAs were fragile. They needed complex multi-term losses, exponential moving averages, or pre-trained encoders to avoid representation collapse. LeWorldModel strips this down to two losses: a next-embedding prediction loss and a regularizer enforcing Gaussian-distributed latent embeddings. That cuts tunable hyperparameters from six to one. At roughly 15 million parameters, the model trains on a single GPU in a few hours. The research team, led by Lucas Maes with Quentin Le Lidec, Damien Scieur, and Randall Balestriero, tested it across 2D and 3D control tasks. It stayed competitive with much larger models. The latent space also encodes genuine physical structure. Probing experiments showed it captures physical quantities, and surprise evaluation confirmed it can flag physically implausible events. LeCun's thesis just got fresh backing. Predicting in latent space beats generating pixels. You don't need to reconstruct every texture to understand physics. A lean model can learn useful world representations without the computational overhead of massive generative systems.