LeCun World Model Progress, Based on Pre-Trained Visual Features, Enables Zero-Sample Planning

Yann LeCun emphasizes that existing LLMs cannot truly understand the world, and promotes data-based dynamic world models; the DINO-WM world model developed by LeCun's team can achieve zero-sample planning by pre-training visual features; DINO-WM performs excellently in complex environments, generalizes effectively to new environmental configurations, and demonstrates a high level of scene understanding.