World Models Need Causality, Not Pretty Pixels — Christopher Manning, Moonlake AI

Why watch

Christopher Manning argues that current generative video models like Genie 3 simulate observations without underlying semantics, making them unsuitable for planning. He advocates for embodied intelligence built on simulation, where AI constructs action-conditioned world models from visual input. His company Moonlake AI builds interactive 3D environments from single photos or videos, enabling object manipulation and interaction. The system researches objects on the web to fill in occluded information (e.g., tea bags in a closed box), using a Claude Code-inspired loop to compare renders against reality and close the sim-to-real gap. The talk covers AI history from Dartmouth 1956 through modern LLMs, targeting AI researchers and engineers working on world models, robotics simulation, or embodied intelligence.