DeepMindthe division of Google focused on artificial intelligence, has announced Genie, “a new paradigm” within generative AIs. And it is not an exaggeration. In recent months we have seen artificial intelligences that create images, music, video and voices from different types of data inputs in an extraordinarily simple way. Now it's the turn of video games.
Genie allows you to create a interactive gaming environmentfrom a single reference image that can be synthetic, real or even a sketch. And also through promptsas is usual in these tools.
This new base model for playable worldsas DeepMind calls it, has been trained from more than 200,000 hours of videos publicly available at Internet . But you are not limited by them, you can work from new images that have not been part of the training. For example, a photograph of their surroundings that the user takes.
Google DeepMind just dropped 'Genie', an AI that can generate interactive video games.
This is a huge deal.
Genie is trained on 200,000 hours of unsupervised public internet gaming videos and can generate video games from a single prompt or image.
But here's what's insane:… pic.twitter.com/C63tDAVEVr
— Rowan Cheung (@rowancheung) February 26, 2024
“What makes Genie unique is its ability to learn precise controls exclusively from Internet videos“, they point out from the company. Genie determines who the main character is and learns how to control him and how he interacts with the virtual world around him in space and time; and he is also able to simulate the object deformation.
DeepMind highlights that during training they have not been used action tags that indicate what action is being performed in a video, something that videos on the Internet do not usually have. Still, the AI not only learns which parts of the image it analyzes are controllable but also infers actions that are consistent with the generated environment.
The Google division has opened a website showing numerous examples of what Genie can do. In them you can see how a single image is enough, which can be created with a text-to-image AI, as can be seen in some of them, for the AI to generate a playable interactive environment. In this case they have used Image2, another generative AI from Google. Other examples are made from real-world photographs and sketches. Genie supports written instructions, but is something DeepMind mentions only in passing.
Genie's training has been focused in platform games and roboticsbut DeepMind assures that “our method is generalshould work for any type of domain and is scalable to increasingly large Internet data sets.”
According to the company, video games are an effective test bed for artificial intelligence, but are limited by the games available. With Genie it is possible to create any world that is necessary, something that DeepMind believes can be “a catalyst for training generalist AI agents of the future”.
Google does not reveal what other uses it will give to this artificial intelligence or Whether it will be available to users in the futureor, either as part of Gemini or independently. Despite the captivating concept, it is obvious that the tool is in an initial phase. But seeing how generative AI has advanced in the last year, we can expect anything in the coming months.