Google DeepMind launches Genie 3, a general-purpose world model that turns text into navigable 720p realms running at 24 frames per second. A decade of simulated-environment research, from game agents to open-ended robotics, pushed the team to build systems that forecast how spaces evolve and how actions reshape them. Real-time interactivity now joins consistency and realism as core capabilities, letting users steer and linger inside generated scenes for minutes without collapse.

Long-horizon coherence arrives as auto-regressive generation battles accumulated error while revisiting locations after elapsed time. Worlds hold physical logic across several minutes, with visual memory stretching back roughly one minute, so structures remain intact when returning to view. Promptable world events deepen control, letting text alter weather or inject characters to test counterfactual choices and agent responses under volatile conditions.

Embodied agents such as SIMA operate inside these spaces to pursue multi-step goals while Genie 3 simulates consequences blind to intent. Current limits include narrow direct action, shaky multi-agent dynamics, imperfect geography, and brief sessions rather than hours. Limited research preview access goes to academics and creators as Google DeepMind refines safety and utility.