Google DeepMind unveils Genie 4 world model in research preview

Genie 4 builds on Genie 3, roughly doubling the maximum simulation duration to 2 minutes, and adds real-time audio plus Gemini-driven direction that breaks user goals into subtasks and checks each rendered frame. Google also pairs it with its video model Gemini Omni for complex scenes, such as maple seeds spinning onto a lawn or a crowded 1980s station platform.

The demos that accompany the release also reveal limits. Beyond the marble bust that fails to roll, the checker-playing Apollo robot demo does not explain how the actual robot control policy is wired into the generation loop. That gap matters because reinforcement learning depends on learning from failure, and a director model that checks for step success may bias environments toward intended outcomes — the body cites the bust's clean landing as an example of that drift.

Google's earlier experiment points to a possible path: combining language and code models for structured physical reasoning with pixel-rendering generators. The industry is already split between real-time video generators like Google and Runway, persistent 3D scene builders like World Labs — which AMD agreed to acquire for about $8.2 billion — and autonomous-driving simulators from Waymo and Rivian.