DeepMind reveals Genie 3, a world model that could be the key to reaching AGI

DeepMind, a subsidiary of Google, has unveiled its latest innovation, Genie 3, a groundbreaking world model that aims to pave the way towards achieving artificial general intelligence (AGI). During a recent press conference, Shlomi Fruchter, a research director at DeepMind, emphasized the model's significance, stating, "Genie 3 is the first real-time interactive general-purpose world model. It transcends the limitations of previous narrow world models, offering versatility across various environments." Genie 3, currently in a research preview and not yet available to the public, enhances the capabilities of its predecessor, Genie 2, which was designed to generate new environments for AI agents. Additionally, it incorporates advancements from DeepMind’s Veo 3, a cutting-edge video generation model renowned for its deep understanding of physics. With just a simple text prompt, Genie 3 is capable of generating several minutes of diverse, interactive 3D environments at a frame rate of 24 frames per second and a resolution of 720p—an impressive upgrade from the mere 10 to 20 seconds it could achieve with Genie 2. A notable feature of Genie 3 is its ability to implement "promptable world events," allowing users to modify the generated world using prompts. One of the model's key strengths is its ability to maintain physical consistency over time, as it remembers previously generated elements—an emergent capability that researchers did not specifically program into the system. Jack Parker-Holder, a research scientist at DeepMind, noted that while Genie 3 has clear implications for fields such as education, generative media, and gaming, its true potential lies in training agents for versatile tasks, which is critical for progressing towards AGI. Fruchter elaborated on the model's architecture, stating, "The model is auto-regressive, generating one frame at a time and referencing prior frames to predict future outcomes. This memory creates coherence in its simulated worlds and fosters an intuitive understanding of physics, akin to human comprehension." This ability to simulate realistic, physically plausible environments over time positions Genie 3 as more than just a generative model; it serves as a robust training ground for general-purpose agents. The model can create an infinite array of worlds for exploration, pushing agents to adapt, learn, and experience challenges in a manner reflective of human learning. Despite its advancements, Genie 3 does face limitations. Currently, the range of actions available to agents is restricted, and while the promptable world events allow for diverse environmental modifications, they are not necessarily executed by the agents. Additionally, the model supports only a few minutes of continuous interaction, whereas extended periods would be more beneficial for comprehensive training. Nevertheless, Genie 3 marks a significant leap in the endeavor to cultivate agents capable of planning, exploration, and learning through trial and error—essential aspects of self-directed, embodied learning that are vital for achieving general intelligence. Parker-Holder remarked on the current state of embodied agents, stating, "We haven’t yet witnessed a Move 37 moment for these agents, a point where they can take creative actions in the real world. However, with Genie 3, we are on the brink of ushering in a new era."