What changes compared with the past
Google DeepMind announced Genie 4 on October 27, a model capable of generating interactive virtual worlds in real time. According to a report by ITmedia AI+, the system succeeds Genie 3 and aims to improve visual rendering and physical simulation: users can move through space and act on the environment, receiving a response generated on the spot rather than a prerecorded video. This is precisely the difference from ordinary video generators: Genie does not play back a ready-made video, but builds the world as it is explored.
The most visible update concerns audio. Genie 4 natively generates ambient sounds such as wind and waves, as well as music and character conversations, which react to situations and user input. It is also possible to modify the world with natural-language instructions: change the weather, insert people or objects, move the camera. DeepMind claims it can simulate phenomena such as aerodynamics and collisions, and identifies the representation of people as the main leap, with figures that move naturally and respond to the environment.
Integration with Gemini and planned applications
The model relies on the Gemini family. According to the source, Gemini Omni contributes to appearance and behaviors consistent with the real world, Gemini 3.8 Flash controls scenes during generation and plans articulated actions, while experiments are underway with Gemini 4 Argon to simulate robots engaged in complex tasks. These are the company's statements, not results independently verified.
The imagined applications range from film to video games, from education to robotics, including the training of AI agents. Distribution will follow multiple channels: first a research preview for selected testers, then access for Google AI Ultra subscribers through Project Genie and for early-access partners through the Gemini API and Google AI Studio. A waitlist is open for Project Genie. Generated content, video and audio, will contain the SynthID watermark to signal its artificial origin.
What remains to be verified
DeepMind itself acknowledges the current limitations. Continuous interaction lasts only a few minutes and, in prolonged sessions, objects and environments can lose consistency in appearance and position. Critical issues also remain in complex physical interactions and audio accuracy. No definite timeline for public release or measurable quality metrics are provided. The picture that emerges is therefore that of a promising but still experimental technology, whose impact will depend on its ability to overcome these constraints and on the actual availability of the various announced access channels.
