Google Deepmind releases a modern version of its “World” model, called Genie 3, capable of generating 3D environments with which users and agents AI can interact in real time. The company also promises that users will be able to interact with worlds much longer than before and that the model will remember where they are when you look away from them.
World models are a kind of AI system that can simulate environment for purposes such as education, entertainment or lend a hand in training robots or AI agents. Thanks to the global models, you give them a quick prompt and generate a space in which you can move like in a video game, but instead of the world it is hand made with 3D resources, all this is generated with AI. This is an area in which Google puts a lot of effort; The company showed Genie 2 in December, which could create interactive worlds based on the image, and builds a world models team led by former cooperation tools for generating Sorai video in OPENAI.
But the models currently have many disadvantages. For example, Genie 2 Worlds was only up to a minute. Recently, I tried the “Interactive Video” company supported by the co -founder of Pixar and I felt like going through the blurred version of Google Street View, in which everything changed and changed in a way that I did not expect when I looked around.
Genie 3 seems to be a significant step forward. Users will be able to generate worlds with monitors, which supports “a few” minutes of continuous interaction, which compared to 10-20 seconds of interaction with Genie 2, According to post on the blog. Google says that Genie 3 can keep the spaces in visual memory for about a minute, which means that if you turn away from something in the world, and then turn to it, things such as paint on the wall or writing on the board will be in the same place. The worlds will also have a resolution of 720p and operate with 24 frames per second.
Deepmind also adds “quick world events” also to Genie 3. By using print, you will be able to do things such as changing weather conditions in the world or adding modern characters.
However, this is probably not a model that you will be able to try yourself. It is introduced as a “limited research preview”, which will be available to the “little cohort of scientists and creators” so that his developers can better understand the risk and how to relieve them according to Google. There are also many restrictions, such as circumscribed ways of interaction of users with generated worlds, and this legible text is “often generated only if it is delivered in the description of the entrance world.” Google claims that this is a “study”, how to bring Genie 3 to “additional testers”.