DeepMind thinks its new Genie 3 world model presents a stepping stone toward AGI

Google DeepMind has introduced Genie 3, an innovative foundation world model poised to train general-purpose AI agents. This breakthrough, as described by research director Shlomi Fruchter, represents a significant advancement towards achieving artificial general intelligence (AGI), which mimics human-like cognitive capabilities. "Genie 3 is the first real-time interactive general purpose world model," Fruchter explained at a recent press briefing. Unlike previous narrow models, Genie 3 is not confined to specific environments; it can create both realistic and imaginative worlds. Currently available only in a research preview, this new model builds upon its predecessor, Genie 2, which specialized in generating new environments for agents, and DeepMind’s latest video generation model, Veo 3, known for its deep understanding of physical principles. With a simple text prompt, Genie 3 can create multiple minutes of interactive 3D environments at 720p resolution and 24 frames per second, a notable improvement over the 10 to 20 seconds of content generated by Genie 2. Additionally, it boasts the feature of "promptable world events," allowing users to alter the generated worlds through prompts. One of Genie 3's most remarkable capabilities is its ability to maintain physical consistency over time, as it retains memory of its prior outputs—an aspect that DeepMind engineers did not explicitly program into the model. Fruchter emphasized the model's potential beyond gaming and educational applications, highlighting its role in training agents for general-purpose tasks, which he believes is crucial for reaching AGI. Jack Parker-Holder, a research scientist on DeepMind’s open-endedness team, pointed out that world models are essential for embodied agents, particularly in simulating real-world scenarios, and Genie 3 is designed to address this challenge. Unlike Veo, Genie 3 does not depend on a hard-coded physics engine. Instead, it learns the dynamics of the world on its own, understanding how objects interact and move by recalling previous generations and reasoning over extended time frames. "The model generates one frame at a time and must refer back to previous outputs to determine subsequent actions," Fruchter told TechCrunch. This memory feature enhances the consistency of the simulated worlds, helping the model develop a grasp of physics, akin to human intuition, such as recognizing when a glass is about to fall. Moreover, Genie 3 has the potential to push AI agents to their limits by enabling them to learn from experience. DeepMind demonstrated this by testing Genie 3 with its Scalable Instructable Multiworld Agent (SIMA), instructing it to achieve goals in a simulated warehouse environment. Parker-Holder stated that the SIMA agent successfully completed tasks like moving towards a bright green trash compactor or a packed red forklift, thanks to the consistency of Genie 3’s simulations. However, Genie 3 is not without its challenges. While it claims to understand physics, some demonstrations, such as a skier down a mountain, did not accurately reflect snow movement. Additionally, the range of actions available to agents remains limited, and complex interactions among multiple independent agents in shared environments are still difficult to model accurately. Currently, Genie 3 supports only a few minutes of continuous interaction, whereas longer sessions would be ideal for thorough training. Despite these limitations, Genie 3 marks a substantial step in teaching agents to move beyond mere reactions to inputs, potentially allowing them to plan, explore, and learn through trial and error—key elements for achieving general intelligence. "We haven’t had a ‘Move 37’ moment for embodied agents yet, where they can take unexpected actions in the real world," Parker-Holder remarked, referring to the iconic move in the game of Go by DeepMind’s AlphaGo. “But with Genie 3, we may be on the brink of a new era in AI development.”