Google DeepMind has announced the release of Genie 3, a groundbreaking general-purpose world model that can generate an unprecedented diversity of interactive environments in real-time. This latest innovation marks a significant milestone in the development of world models, which are AI systems capable of simulating aspects of the world to enable agents to predict its evolution and their actions’ effects.
Genie 3’s capabilities include modeling physical properties, simulating natural phenomena, animating fictional scenarios, exploring historical settings, and more. The model achieves real-time interactivity at a frame rate of 24 frames per second, with visual consistency extending for several minutes at a resolution of 720p.
To test the compatibility of Genie 3 created worlds with future agent training, researchers generated worlds for a recent version of their SIMA agent. This enables the execution of longer sequences of actions and achieving more complex goals.
However, Genie 3 also comes with limitations, including limited action space, interaction and simulation of other agents, accurate representation of real-world locations, text rendering, and limited interaction duration.
While acknowledging its current limitations, Google DeepMind is committed to developing foundational technologies in a way that amplifies human creativity while limiting unintended impacts. The company is exploring how to make Genie 3 available to additional testers in the future, with the aim of harnessing this technology for education and training purposes.
Source: https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models