Google DeepMind Launches Genie 4 World Model
Google DeepMind introduces Genie 4, an interactive world model with real-time scene changes, sound and longer simulations.
AI-generated video is evolving into environments that respond to actions rather than simply playing back a scene. Google DeepMind's Genie 4 is another step toward that goal.
What happened
Google DeepMind introduced Genie 4 as a research preview. The world model generates interactive, increasingly photorealistic environments with synchronised audio, and allows users to change scenes through prompts while simulations are running.
Its coherent sequences can last around two minutes, approximately double the previous model's duration. Integration with Gemini enables a model to plan steps and direct the simulated scene. In an early robotics experiment, a Gemini Robotics system broke down tasks that Genie simulated.
Access is initially restricted to selected testers, with broader access planned through Google's experimental products and development tools. Genie 4 still struggles with physical consistency over time, drifting objects, readable text and reliable audio.
Why it matters
World models could be useful for building interactive entertainment, prototyping creative ideas and testing embodied AI systems. Training and evaluation in simulated environments may reduce the expense and risk of collecting every experience in the real world.
But visually convincing environments are not necessarily physically accurate ones. For robotics, incorrect friction, contact or object behaviour can teach the wrong lessons.
The bigger picture
World models are becoming a competitive layer of AI infrastructure, connecting video generation, simulation and decision-making. Genie 4 strengthens that direction, although it remains a research product rather than a proven replacement for physics engines or real-world robot testing.