Summary
Peter Yang gets early hands-on time with Google’s Genie 4, a world model that generates interactive scenes frame by frame, and builds five worlds live. He describes world models as doing for frames what language models do for words: predicting what comes next based on user input, such as pressing W to move forward.
Yang highlights three new capabilities in Genie 4: generated sound, including speech, ambient noise and music; live changes to a scene by typing prompts such as a shooting star; and Gemini working behind the scenes to plan multi-step events, shown with a robot arm completing a lab task.
The five scenes include a walk through San Francisco interrupted by a kaiju attack, ordering at a 1980s Tokyo ramen stand where the chef talks back, storming a castle on a dragon, a jeep safari with lions and elephants, and defending New York from aliens. Worlds generate in roughly 10 seconds. Yang is candid about flaws, including stacked ramen bowls, odd dragon sounds, a jeep that does not drive, and limited resolution. He closes with thoughts on what world models could mean for film and gaming.
📺 Source: Peter Yang · Published October 28, 2026
🏷️ Format: Showcase
