Peter Yang demonstrates Google’s Genie 4 in a 13-minute early-access video, creating five interactive scenes and keeping their failures visible.
Generate the next view from an action
- Yang describes a world model as predicting the next frame from the current scene and user input.
- His demo adds generated audio, live prompt changes, and what he describes as Gemini-assisted event planning.
- He reports roughly ten seconds to initialize the first scene; the video supplies no systematic speed test.
Five scenes, uneven control
- San Francisco: a peaceful street gains a retriever, warmer lighting, and a kaiju attack while the tourist keeps holding a phone.
- Tokyo ramen stand: the chef speaks and serves food, but the requested eating action fails and bowls stack oddly.
- Dragon and castle: spectacle appears quickly, with strange vocal sounds and an intermittently disappearing spaceship.
- Safari: animals appear, but the Jeep does not seem to drive and antelope emerge abruptly.
- Alien attack: the scene looks like a shooter, but firing does not reliably remove the aliens.
What the demo supports
- Interactive scene generation can be entertaining even when actions only partly follow the prompt.
- Yang sees potential for effects and personal experiences; those are forecasts rather than demonstrated production savings.
- A reliable game still needs responsive rules, progression, and a compelling loop beyond generated imagery.
Caveat: this is a creator-presented early-access demonstration, with visible continuity and control failures. It supplies no independent performance or long-term world-consistency benchmark.
“You need to have a core loop. You need to have progression systems.” — Peter Yang