World models need causality, not pretty pixels — Christopher Manning on embodied simulation

AI Engineer · youtube · 2026-09-24

Stanford professor Christopher Manning (now at Moonlake AI) tours AI history — from Dartmouth 1956 to a 2007 Google LM trained on 2T tokens — and argues the next step is embodied intelligence via simulation. Generative video like Genie 3 simulates observations without semantics and can't support planning; Moonlake instead builds action-conditioned world models in code from a single photo, researching objects on the web to fill invisible details, with a Claude Code-inspired loop comparing renders to reality to close the sim-to-real gap and replace 10,000 hours of teleoperation.

Original post →

More from Embodied

Embodied channel →