DeepMind Unveils Genie: An AI Model Crafting Games from Prompts

On February 26, Google’s subsidiary DeepMind introduced Genie, a generative AI model capable of creating games. Trained on over 200,000 hours of gameplay videos, the neural network can transform any image or text into a 2D platformer.

The name Genie is a play on words. It stands for Generative Interactive Environments and is a nod to the Genie character from the Aladdin cartoon.

Rather than adding inductive biases, we focus on scale. We use a dataset of >200k hours of videos from 2D platformers and train an 11B world model. In an unsupervised way, Genie learns diverse latent actions that control characters in a consistent manner. pic.twitter.com/71a3iuAGZA

— Tim Rocktäschel (@_rockt) February 26, 2024

“Recently, we have seen the emergence of generative AI models capable of creating diverse content from text, images, and even video. Today, we introduce a new paradigm—generative interactive environments: Genie,” said DeepMind representatives.

According to the creators, Genie is trained on a large set of publicly available materials from the internet. The team focused on videos related to 2D platformers and robotics. It is noted that its algorithm can operate in any domain and scale to other data sets.

The developers believe that the advent of Genie is also significant for training AI agents. Their research has shown that the gaming environment can be an effective testing ground. Thanks to the new neural network, they can learn in an infinite number of worlds created every day.

Earlier in February, on the 21st, Google unveiled the open AI model Gemma, which, according to test results, surpasses neural networks from Meta and Mistral.