DeepMind opens Project Genie to U.S. Google AI Ultra users, seeks real-world feedback on interactive world models

DeepMind has initiated a limited rollout of Project Genie, an experimental system that converts text prompts and images into short, navigable virtual environments, to a narrow set of U.S. Google AI Ultra subscribers. The prototype stitches together DeepMind’s latest world model work—built on the Genie 3 research backbone—with an image generator and other multimodal components to produce playable spaces in seconds. Access is intentionally restricted and sessions are capped, reflecting heavy compute demands and the team’s desire to gather focused user data rather than offer an open consumer product. In practice the tool handles stylized, whimsical prompts reliably, producing charming, game-like environments, while struggling to reproduce photorealism or cinematic fidelity. Users can begin with an automatically generated image or upload a photo, then let the system expand that input into an explorable scene; the experience mixes genuine emergent interactivity with frequent glitches such as collision failures and errant navigation. DeepMind has already embedded safety filters that block explicit content and attempts to recreate recognizable copyrighted characters, narrowing some creative use cases but reducing legal and ethical exposure. The project’s auto-regressive architecture helps maintain consistency across short revisits to generated areas, though repeatability is imperfect. Input handling and control responsiveness remain work-in-progress: navigation keys can feel unsteady and object behaviors occasionally defy physical intuition. Limiting sessions to roughly one minute is a deliberate trade-off—the firm is balancing per-user computational costs with the practical need to scale testing. This staged, feedback-driven release strategy positions DeepMind to iterate quickly, harvest human reactions for model improvement, and explore commercial pathways that begin with entertainment and simulation-driven training for embodied agents. Competitors and new entrants are racing similar capabilities, so the preview also functions as a public benchmark and a data-acquisition channel. For now, Project Genie is a research-first demonstration: it reveals the potential of world models to make interactive content from simple inputs while highlighting gaps in realism, control, and infrastructure needed for wider deployment.