DeepMind's Genie 4 turns a single image or prompt into a steerable, audio-enabled world
Google DeepMind has announced Genie 4, a world model that generates explorable environments with native audio and lets users rewrite the scene in natural language while it runs. It is in research preview for trusted testers, with early access planned via Project Genie and the Gemini API.
Google DeepMind has introduced Genie 4, the latest generation of its world-simulation models. Where earlier versions produced short interactive clips, Genie 4 generates coherent environments that users can navigate and modify in real time by typing instructions, changing the weather, moving the camera, or adding a person or object mid-session.
Gemini sits in the loop. Between user inputs it monitors the scene to keep it consistent, and it can decompose a goal into steps, verify each one and re-plan if the user intervenes. Physics and character motion have improved, with more plausible aerodynamics, collisions and human movement, and the model generates ambient sound and interactive dialogue natively alongside the visuals.
DeepMind is pitching the model at four audiences: filmmakers who want rapid previsualisation, journalists and historians who can reconstruct an environment from a single archival photograph or visualise a flood or wildfire, game designers prototyping mechanics from text, and enterprises generating training scenarios for robots and agents.
Limitations are acknowledged candidly. Object permanence and visual detail can drift over long sessions or after many edits, and text rendering, spatial accuracy and audio fidelity are still active research problems. Genie 4 is currently a research preview for trusted testers, with a waitlist through Project Genie for Google AI Ultra subscribers and early-access partners via the Gemini API and AI Studio.
Why it matters
World models are becoming the training ground for embodied AI. If simulated environments are cheap, varied and controllable, robots and agents can be trained on scenarios that would be dangerous or impossible to stage physically, which shortens the path from lab to deployment.
Indicops view
The near-term applications we see for Indian public bodies are in disaster preparedness and training: visualising flood scenarios for district administrations, or generating drill environments for emergency responders. These are not production workloads yet, but they are worth piloting now so that institutional knowledge exists when the tools mature.
More in Research
Back to the newsroomAnthropic study finds AI agent teams do worse when each user brings their own agent
Anthropic's alignment team tested frontier models in four multi-user settings and found group outcomes were worse when each person delegated to a personal agent than when one coordinator served everyone. Three of the environments are released as the MAMUBench benchmark.
DeepMind says AI agents found two weaknesses in legacy 1024-bit RSA
Google DeepMind says AI models working under human direction found two weaknesses in legacy 1024-bit RSA, covering signature forgery and cheaper batch factoring of related moduli. Keys of 2048 bits or more remain secure against this specific threat, according to the reporting.