A new chapter in digital storytelling may be unfolding, as AI lab Odyssey has unveiled the first public research preview of what it calls “interactive video” a format where viewers don’t just watch, but also interact with AI-generated content in real-time.

Unlike traditional video or even pre-rendered game engines, this new experience is powered entirely by a world model, a type of AI architecture that predicts how a scene should evolve based on past states and incoming user actions. Whether using a keyboard, touchscreen, or controller, users can influence the video stream as it unfolds frame by frame.

The underlying model generates realistic video frames every 40 milliseconds, enabling interactivity that feels responsive and fluid. Odyssey’s infrastructure, powered by clusters of H100 GPUs across the U.S. and Europe, can deliver video at up to 30 frames per second. From input to visual output, actions taken by the user are reflected on-screen with minimal delay demonstrating the early capabilities of a real-time AI video system.

While the experience is described by the team as “glitchy” and “raw,” it marks a meaningful technical milestone. The model differs significantly from traditional video models, which generate entire clips in advance and are inherently non-interactive. Instead, Odyssey’s world model continuously predicts the next visual state, enabling open-ended interactions that evolve in response to user input.

To improve short-term stability, the team employed a post-training approach, training the model on a smaller set of densely covered video environments. This reduces generality but allows for longer, more coherent video generation during interaction. A more generalized, next-generation world model is already in development, with early outputs showing improved realism and broader action understanding.

What sets this apart from previous research that often focused on constrained video game environments like Minecraft or Quake is that Odyssey’s approach learns from real-world video. This shift introduces more complexity but opens the door to more lifelike visuals, nuanced motion, and an unbounded action space beyond pre-scripted logic.

Odyssey sees this as the beginning of a new medium one that could eventually reshape how we think about entertainment, education, advertising, and training. While still in its early stages, the project shows potential to evolve beyond static narratives toward experiences that adapt and respond in real time.

The team is transparent about the ongoing challenges, especially in stabilizing long-form autoregressive generation and modeling open-ended human actions. But the research preview offers a glimpse of what’s possible when storytelling meets real-time generative intelligence.

Odyssey invites the public to try the experience and welcomes feedback as it continues to push the boundaries of interactive AI.