We’ve seen AI generate video, but DeepMind Genie just made it interactive. In this technical briefing, we dive into the "Foundation World Models" that are turning static pixels into 720p playable worlds. From reverse-engineering controllers to training the next generation of robots, this is the future of embodied AI.

Key takeaways:

• The LAM Breakthrough: How the Latent Action Model learned to "play" just by watching YouTube.

• Physics-Compliant Worlds: Why Genie 3 understands cause and effect better than typical video generators.

• The Nintendo Problem: Why Google had to filter out "accidental" recreations of Mario 64.

• Embodied AI: Why the real "players" of these games aren't humans, but robots like the Agibot GO1.

Chapters:

(00:00)Introduction: The Shift to Interactive AI

(00:21)Project Genie vs. Sora: Gameplay over Video

(01:04)The Brain: Latent Action Models (LAM) Explained

(01:39)Reverse-Engineering the Controller

(02:09)Spatio-Temporal Tokenizers: Real-Time Dreams

(02:45)The Real Customer: Training Robots, Not Humans

(03:53)Reality Check: Computation & Hallucinations

(04:25)The Nintendo Knock-off Problem & IP Filters

(05:12)Final Thoughts: When Simulation Becomes Reality