Google’s AI research division, DeepMind, has announced a new generative model called Genie, which stands for Generative Interactive Environments. This model is capable of creating fully playable, interactive 2D platformer games from a single prompt, such as a photograph, a sketch, or a text description. The researchers at DeepMind refer to this creation process as “vibe-coding.”
The announcement showcased Genie’s ability to take a static image and generate a complete game world based on it. This technology represents the first generative interactive environment model of its kind, according to its creators. The 11-billion parameter model works by generating the game environment one frame at a time, allowing for real-time interaction from a user.
Training on Unlabeled Video Data
To achieve this, the Genie model was trained on a vast dataset consisting of over 200,000 hours of video from publicly available 2D platformer games. Significantly, this training data was unlabeled, meaning the AI was not explicitly instructed on game mechanics, character controls, or which parts of the video were controllable. Through this unsupervised learning process, Genie independently learned to infer actions like walking, running, and jumping.
A Foundational World Model
By analyzing the video content, Genie identifies the main character and learns a latent action space, which allows a user to control the character within the newly generated environment. The researchers describe Genie as a foundational world model that can be prompted to create an unlimited variety of playable worlds. At the time of its announcement, Genie was presented as a research preview and has not been released to the public.
Concise Cyber 