DeepMind unveils Genie 4 world model promising more realistic physics and humans

Google DeepMind has introduced Genie 4, an updated world model that generates explorable virtual environments frame by frame in real time from user prompts. The system can create experiences resembling video games without pre-rendered environments or individually programmed interactive elements. DeepMind positions world models as a component of its work toward artificial general intelligence and says they could provide training environments for physical robots and support rapid creative prototyping. Genie 4 uses Gemini 3.8 Flash as a guiding agent that breaks down goals and evaluates generated frames, helping simulations retain context as users move through different environments. DeepMind says the model improves on Genie 3 in several areas, including more useful simulations, but acknowledges that generated worlds remain consistent for only a couple of minutes before losing track of their initial state. Users can alter a simulation while it is running by changing weather, adding elements, or moving the camera; Gemini helps steer the generated world after each update. Google DeepMind says Genie 4’s interactive generation does not require environments to be created in advance, though access may depend on paying Google.