AI "Genie" Unveiled by Google DeepMind to Instantly Create Playable Games

The experimental model, which has been trained on more than 200,000 hours of gaming videos, can transform any image or concept into a 2D platformer.
Genie, short for Generative Interactive Environments, was developed in partnership with Google and the University of British Columbia. With just one image, it can generate side-scrolling 2D platformers such as Contra and Super Mario Brothers based on human suggestions.

According to Google DeepMind, "generative AI has emerged in the last few years, with models capable of generating novel and creative content via language, images, and even videos." "We are pleased to present Genie, a new paradigm for generative AI and generative interactive environments."


Thanks to what Google researchers refer to as a latent action model that infers the actions between video frames, a video tokenizer that turns raw video frames into discrete tokens, and a dynamic model that determines the next frame, Genie can create interactive, playable environments from a single image prompt.

"We prioritize scale over inductive biases," Tim Rocktäschel, a Google DeepMind developer, stated on Twitter. "We train an 11B world model using a dataset comprising over 200k hours of videos from 2D platformers. Then, unsupervised, Genie learns diverse latent actions that control characters in a consistent manner."


Rocktäschel went on, "Genie can also turn other kinds of media into games." Genie can be asked to create a range of action-controllable virtual worlds from a number of inputs in the accompanying Google DeepMind research paper.

According to Rocktäsche, "Our model can turn any image into a playable 2D world." "Sketches and other human-designed creations, like the exquisite artwork of Seneca and Caspian, two of the youngest artists in history, can come to life with Genie."


"We also show that we can learn an action controllable simulator there as well by training a Genie on robotics data (RT-1) without actions," the speaker added. "We believe that this is a positive step toward AGI general world models."

Artificial general intelligence (AGI), also referred to as the singularity, is the ability of an AI to comprehend and apply learned information to a variety of tasks in a manner similar to that of a human.
The dataset from Genie, according to Google DeepMind, was created by sifting through publicly accessible online videos, particularly ones with titles like "dpeedrun" or "playthrough," and removing terms like "movie" or "unboxing."

According to Google DeepMind, advances in AI hardware, software, and datasets have made it possible to produce "crisp and aesthetically pleasing" visuals as well as coherent, conversational English.

The researchers added, "When choosing keywords, we manually spot checked results to make sure they typically produced gameplay videos for 2D platformers that aren't outnumbered by other types of videos that also happen to share similar keywords."


Google DeepMind stated, "With Genie, our future AI agents can be trained in a never-ending curriculum of new, generated worlds." "This is just scratching the surface of what may be possible in the future; in our paper, we have a proof of concept that the latent actions learned by Genie can transfer to real human-designed environments."

Large investments in generative AI have been made by major giants like Google, Microsoft, and Amazon, in large part because of the release of OpenAI's GPT-4 last year. Google, which rebranded from Google Bard, announced earlier this month the release of a subscription-based version of its Gemini AI model.