Yann LeCun left Meta, raised Europe’s biggest seed and made an AI breakthrough (again!) yannlecun amilabs.xyz 👏🏼 AI might not need to get bigger to get better. This new paper from Meta’s former Chief AI Scientist and the Godfather of AI tackles a hidden problem in world models: latent space collapse. When that happens, different situations look the same to the model, so it stops truly understanding the world. Most solutions add complexity: extra networks, multiple losses, pretrained encoders… This one does the opposite. It simplifies everything: just predict the next state + enforce a Gaussian latent space. The result? • ~15M parameters • Trains in hours • 48x faster planning • Learns real structure (not just patterns) It even reacts more to physically impossible events than superficial changes. That’s a big signal: the model is learning how the world actually works. It’s not perfect, still struggles in complex visual scenes. But it shows something important: Progress i...