Google DeepMind has unveiled “Project Genie,” a pioneering artificial intelligence initiative that allows users to create and interact with dynamic virtual worlds in real-time, marking a significant step towards Artificial General Intelligence (AGI).
At the core of Project Genie is “Genie 3,” which Google describes as a “World Model.” This differs from standard AI models that generate videos because it simulates the environment’s dynamics, allowing the virtual world to respond to user movements.
The system aims to help AI understand real-world physics and interactions, similar to human comprehension. This development is crucial for DeepMind’s broader goal of creating AGI.
Currently an experimental research prototype, Project Genie enables users to define virtual environments, choose avatars, and generate a “World Sketching” preview using a tool called “Nano Banana Pro.” Users can then refine this sketch before entering the world.
Upon creation, the system generates a 720p virtual world where users can interact for up to 60 seconds. The experience runs at 20 to 24 frames per second, a notable achievement for real-time generation.
As users navigate, Genie 3 generates the path ahead in real-time. Google claims the physics engine, including collision and impact consistency, represents a “breakthrough.” Users can also rotate the camera and download videos of their interactions.
Additional features include “Remix Worlds,” which allows users to modify existing virtual environments, and a gallery for exploring new ideas.
Google acknowledges that the generated worlds may not always be 100% realistic or perfectly align with user commands. Interaction may also experience some latency or control difficulties, given its experimental phase.
The company plans to enhance Project Genie with “promptable events,” allowing the worlds to change based on specific user commands. This functionality could significantly benefit fields such as robotics, where it could simulate training scenarios, and animation.
Access to Project Genie is currently limited to Google AI Ultra members aged 18 and over in the United States. Availability for other regions has not yet been announced.