Odyssey Introduces AI-Powered Video Worlds You Control in Real Time

Odyssey, a London-based AI lab, released a research preview for a model that turns standard video footage into fully interactive environments. This pilot focuses on applications in film and game production. During development, the team realized that the underlying technology might give rise to a brand-new medium for interactive entertainment. The research preview is available to industry professionals and academic collaborators exploring immersive media.

Interactive video generated by Odyssey’s system reacts instantly to user inputs. Viewers can pilot the scene with a keyboard, smartphone, or game controller. Voice-driven commands will arrive in forthcoming updates. Odyssey refers to this initial experience as “an early version of the Holodeck.” Early demos are shared with selected partners for testing.

The AI back end achieves updates every 40 milliseconds, producing smooth, realistic views. Each time a button press or gesture is detected, the next frame appears almost immediately on screen. The system creates a compelling illusion of control and presence.

Odyssey commented, “The experience today feels like exploring a glitchy dream—raw, unstable, but undeniably new.” The lab points out that the visuals remain far from AAA-game quality at this stage, leaving room for refinement as research progresses.

A core distinction between this system and conventional video games or CGI lies in its world model approach. Traditional engines pre-render scenes based on scripted events. Odyssey’s model predicts each new frame on the fly, using both the current video state and any recent user actions.

The process bears resemblance to large language models that anticipate the next word in a sentence. In this case, the AI handles vast arrays of pixels instead of text tokens, making the task orders of magnitude more demanding in terms of computation and memory.

Odyssey defines a world model as “an action-conditioned dynamics model.” Each user interaction feeds into a calculation that accounts for prior states, the specific command, and the sequence of frames that led up to that moment. The result forms the next visual output in real time.

The interplay between AI predictions and user direction yields an experience that resists formulaic scripting. There is no rigid logic mapping each input to a fixed outcome. This unpredictability gives the interactive scenes an organic quality, but it also introduces new risks for maintaining coherent visuals.

One major challenge with frame-by-frame generation is drift, a tendency for small errors to compound over time. AI researchers have studied drift for years. Without intervention, brief glitches can grow until the scene breaks down. Odyssey confronts this instability by constraining its training regime.

The lab applies a method it calls a “narrow distribution model.” Initially, the AI trains on a broad collection of general video footage. Then it undergoes fine-tuning on a smaller, more focused set of environments. This strategy limits the range of scenarios the system attempts but yields greater visual stability.

The company reports “fast progress” on its next-generation model, which shows “a richer range of pixels, dynamics, and actions.” Early tests indicate crisper detail in textures, more believable movement, and interactions that better reflect user input.

Operating this AI-driven setup in real time carries a nontrivial cost. Infrastructure expenditures run between £0.80 and £1.60 (1–2) per user-hour. The platform relies on clusters of H100 GPUs distributed across facilities in the US and the EU.

Those numbers may exceed typical streaming fees, yet they represent savings compared with budgets for conventional film shoots or game development. Odyssey expects these expenses to fall sharply as next-generation architectures emerge and efficiency improves.

Each era of media innovation has introduced fresh narrative styles. Early cave paintings paved the way for manuscripts. The invention of the printing press enabled mass readership. Photographic realism inspired audio dramas on the radio. Motion pictures allowed scenes to come alive. Video games then brought immersive, interactive storytelling.

Should AI-generated interactive video reach production quality, applications could span far beyond entertainment. Medical training scenarios might allow students to practice surgical techniques under simulated conditions. Virtual tourism could enable guided exploration of remote landmarks. Brands may deploy dynamic advertisements that adapt in real time to individual viewer actions.

The research preview remains an early proof of concept. Users testing the model will likely encounter visual artifacts, unexpected behaviors, and limits in scene variety. Odyssey considers feedback from this stage crucial for refining its algorithms and progressing toward a more polished, scalable system.