Google’s Genie: An AI Platform for Video Games

Exploring Google's Genie AI platform and its impact on video games

An AI tool called Google's Genie can create interactive 2D video games based just on an image cue or text explanation. 200,000 hours of unlabeled video footage, mostly from 2D platformer games, were used to train a "world model" that is used in this system. With just a cue or image, Google's Genie can create video games, in contrast to standard AI models that need labeled data and clear instructions. Instead, Genie learns by watching the activities and interactions in these movies.

Google's Genie is not just a passive generator of video games, but also an active player and controller of them. It can not only create the game environment but also control a playable character and determine its movements. It can also interact with the game objects, such as enemies, obstacles, or items, and respond to the game events, such as collisions, rewards, or penalties. Google's Genie can also generate different levels of difficulty and complexity for the games, depending on the user's preference and skill.

Google's Genie works by using three core components: a video tokenizer, a latent action model, and a dynamics model. These components work together to process the video data, identify the actions, and generate the video games.

 Video tokenizer:

This component breaks down the large video data into smaller units called "tokens". These tokens are the fundamental building blocks for Google's Genie's understanding of the visual world. They represent the basic elements of a video game, such as the background, the character, the enemies, the items, and the effects.

Latent action model:

This component analyzes the transitions between consecutive frames in the videos. It identifies eight fundamental actions that are essential for playing 2D platformer games, such as jumping, running, shooting, or interacting. These actions are the "spices" that add flavor and variety to video games.

Dynamics model:

This component predicts the next frame in the video sequence and then generates subsequent video results. It continuously does this until a conclusion is reached. This component is the "chef" that brings everything together and creates the video games.

To use Google's Genie, the user can either provide an image prompt or a text description of the desired video game. For example, the user can provide an image of a forest scene or a text description of "a game where you have to collect coins and avoid enemies". Google's Genie will then use its world model to generate a video game that matches the prompt or description. The user can then play the game and provide feedback to Google's Genie, which will use it to improve its performance and generate new games.