Odyssey-3 is a new generative world model that you can try for free

You can now try generating and exploring interactive AI worlds from text prompts.

AI company Odyssey has launched Odyssey-3, a new generative world model available through a public research preview. The model can create interactive virtual environments in real time from text prompts, allowing users to explore and interact with them from different perspectives. Odyssey-3 is designed to simulate physical processes, predict environmental changes based on actions, and serve as a foundation for controlling various systems, including robotic arms, drones, and characters in video games. A free online demo, Odyssey-3 Flash, showcases the model's capabilities. The base Odyssey-3 model has 14 billion parameters, with a Pro version supporting higher resolutions. While Odyssey claims top scores on physics benchmarks like Physics-IQ Verified and WorldMark, these results have some caveats regarding testing methodology. The company plans to use Odyssey-3 for training AI agents, potentially introducing a new paradigm where agents learn from their own actions within generated environments.

Odyssey-3 represents a significant step in the development of world models, aiming for a unified approach that can control diverse physical and virtual systems. By demonstrating its utility across robotics, autonomous driving, and video games with specialized controllers, the company signals a future where a single underlying model can adapt to vastly different tasks. This has implications for the efficiency of AI development, potentially reducing the need for task-specific models and accelerating the deployment of AI in complex, real-world applications.

The public research preview of Odyssey-3, along with its benchmark claims, enters a competitive landscape where companies like Google DeepMind (with Genie 3) and Fei-Fei Li's World Labs are also developing similar technologies. Odyssey's emphasis on real-time interaction and its potential for novel AI training paradigms suggest a growing industry focus on creating more intuitive and dynamic AI learning environments. The success of these world models will hinge on their ability to accurately simulate physics and generalize across a wide range of control tasks.