Google DeepMind has unveiled Genie 3, its latest foundation world model designed to train general-purpose AI agents towards achieving artificial general intelligence (AGI). The model can generate realistic and imaginary worlds, interact with them in real-time, and simulate complex physics.
Genie 3 surpasses its predecessor Genie 2 by producing interactive 3D environments at 720p resolution for up to multiple minutes, compared to Genie 2’s 10-20 seconds. The model also features promptable world events, allowing users to change the generated environment with a single text prompt.
One of the key capabilities of Genie 3 is its ability to maintain physical consistency over time, thanks to its memory-based architecture. This enables the model to develop a grasp of physics, similar to humans understanding how objects move and interact in the real world.
Genie 3’s simulations have been tested with DeepMind’s generalist Scalable Instructable Multiworld Agent (SIMA), which achieved goals set by the researchers. However, the model still has limitations, including limited range of actions an agent can take and difficulty modeling complex interactions between multiple independent agents.
Despite these limitations, Genie 3 represents a significant step forward in teaching AI agents to learn and improve through self-driven exploration and trial-and-error processes. This capability is crucial for achieving AGI, as it enables agents to plan, explore, and seek out uncertainty – essential qualities of human-like intelligence.
Source: https://techcrunch.com/2025/08/05/deepmind-thinks-genie-3-world-model-presents-stepping-stone-towards-agi