Odyssey is trying to turn video from something people only watch into something they can also explore. The startup, founded by self-driving pioneers Oliver Cameron and Jeff Hawke, has developed an AI model that streams generated video while users interact with it through basic controls.
The result is closer to moving through a 3D-rendered video game than watching a fixed clip. It is available on the web as an early demo, and Odyssey describes the approach as interactive video powered by a new world model.
How Odyssey’s interactive video works
The model generates and streams video frames every 40 milliseconds. Instead of relying on a game engine, Odyssey says the system predicts what should happen next based on the current state of the world, an incoming action, and a history of earlier states and actions.
In plain terms, the model is not just making a single video file. It is continuously producing the next visual moment in response to what the user does. That is why viewers can explore areas inside the generated scene rather than simply play, pause, or rewind.
Odyssey says the model demonstrates several capabilities: generating pixels that feel realistic, maintaining spatial consistency, learning actions from video, and producing coherent video streams for 5 minutes or more. Those goals matter because an interactive video system has to do more than create attractive frames. It must keep enough continuity for the user to feel that the scene still makes sense as they move through it.
Why world models are attracting major AI interest
Odyssey is not alone in pursuing world models. The source article names DeepMind, Fei-Fei Li’s World Labs, Microsoft, and Decart among the companies and researchers working in this area.
The larger idea is that world models could eventually support interactive media, including games and movies. They could also run realistic simulations, such as training environments for robots.
That makes the technology important beyond a single demo. A normal video shows a fixed sequence of images. A world model aims to represent enough about a scene and its behavior that it can generate what happens next when a user or system takes an action.
Odyssey’s demo points toward that ambition. A user can move through a generated environment, and the model keeps streaming new frames in real time. Even when the output is imperfect, the interaction changes the role of video from a finished object into a responsive environment.
The demo shows both promise and limits
Odyssey acknowledges that the demo is still rough around the edges. The generated environments can appear blurry and distorted. They can also be unstable, with layouts that do not always remain consistent.
That instability is important. If a user walks forward for a while or turns around, the surroundings might suddenly look different. For interactive media, that kind of inconsistency can break the feeling that the user is navigating one persistent place.
Still, the performance details show why the demo is notable. Odyssey says the model can currently stream video at up to 30 frames per second from clusters of Nvidia H100 GPUs. The company puts the cost at $1 to $2 per user-hour.
Odyssey says it is working to improve the model quickly. The company says it is researching richer world representations that capture dynamics more faithfully, while increasing temporal stability and persistent state. It is also expanding the action space from motion to world interaction and learning open actions from large-scale video.
Why creatives are watching closely
Interactive video has obvious implications for entertainment, ads, education, training, travel, and other areas where video is already used. Odyssey argues that stories could be generated and explored on demand, without the constraints and costs of traditional production.
That vision is also why the technology is controversial. The source article notes that creatives have mixed feelings about AI tools in this space.
A recent Wired investigation found that game studios such as Activision Blizzard, which has laid off scores of workers, are using AI to cut corners and combat attrition. A 2024 study commissioned by the Animation Guild, a union representing Hollywood animators and cartoonists, estimated that over 100,000 U.S.-based film, television, and animation jobs will be disrupted by AI in the coming months.
Odyssey’s stated position is that it wants to collaborate with creative professionals, not replace them. That distinction will be tested by how the tools are built and used. If interactive video becomes part of production workflows, creative control, editing, and authorship will matter as much as raw generation quality.
Odyssey’s data and creator-tool strategy
Odyssey is taking a different route from many AI labs working on world models. The company designed a 360-degree, backpack-mounted camera system to capture real-world landscapes. Odyssey believes that this can provide a stronger basis for high-quality models than relying only on publicly available data.
The company has also discussed tools for creators. Last December, Odyssey said it was working on software that would allow creators to load scenes generated by its models into Unreal Engine, Blender, and Adobe After Effects so the scenes could be hand-edited.
That matters because a generated world is only one part of a professional creative pipeline. If artists can bring AI-generated scenes into familiar software, they may be able to adjust, refine, and direct the output instead of treating it as a finished result.
Odyssey has raised $27 million from investors, including EQT Ventures, GV, and Air Street Capital. Ed Catmull, one of the co-founders of Pixar and former president of Walt Disney Animation Studios, is on the startup’s board of directors.
The early demo does not yet prove that interactive video will replace today’s production methods. It does show a working direction: video that responds to user input, generated moment by moment, with world models at the center. For Odyssey, the next challenge is turning that technical concept into scenes that remain stable, editable, and useful enough for creators and audiences.