Google’s AI research wing, Deepmind, has revealed an AI model that can generate entire playable 3D worlds and entirely new AAA-style games on the fly. The new model is Genie 2 and is a direct successor to Genie — a similar model released in 2024 that could generate an interactive, real-time scene from a single reference image and accompanying text description.
According to Deepmind’s announcement, Genie 2 is a “world model” that can “simulate virtual worlds, including the consequences of taking any action (e.g. jump, swim, etc.).” As expected, Google is rather secretive about the model’s training data, only stating that “it was trained on a large-scale video dataset and, like other generative models, demonstrates various emergent capabilities at scale, such as object interactions, complex character animation, physics, and the ability to model and thus predict the behaviour of other agents.”
The announcement page also includes several demos of the model-generating worlds in an environment resembling AAA games. The character animations and overall world generation seem quite coherent.
Compared to similar models such as Decart’s Minecraft simulator Oasis, the resolution is quite good, and the model does a good job of remembering parts of a scene it created that aren’t in view and rendering them correctly when they become visible again. Genie 2 can generate consistent worlds for up to a minute, with most examples lasting 10-20s.
That said, it doesn’t mean developers can start creating entire games on Genie 2 right away. If they did, the model would constantly erase their progress, and the resulting “game” wouldn’t be much more than the player running around in a real-time AI-generated world—that too for only about a minute as things currently stand. This is why Deepmind is marketing the model as more of a prototyping tool for creating interactive experiences and AI agents.
However, one big issue that Google might be facing is copyright issues. Deepmind, as part of Google, has unrestricted access to YouTube, and while Google’s terms of service do imply that it has permission to use YouTube videos to train its AI models, just how much of the AI-generated world will resemble an actual game and infringe its IP rights in the process remains to be seen.
In the News: Black Basta ransomware operators refine social engineering tactics
