Odyssey drops real-time interactive video powered by world models

Imagine watching a video where pressing an arrow key instantly steers the camera, reshaping the entire scene on the fly. It is not an Unreal Engine render or a branching Netflix special. The entire visual feed is being dreamed up by artificial intelligence frame by frame, in real time.

London-based AI lab Odyssey has launched a public research preview of what it calls interactive video, pitching it as a brand-new medium for storytelling powered by world models.

A glitchy dream running at 40 milliseconds

Odyssey was founded by self-driving car veterans Oliver Cameron and Jeff Hawke. The duo originally set out to build world models to streamline visual effects and game production, but they ended up stumbling into a completely different format.

Their current demo streams up to 30 frames per second using Nvidia H100 GPUs. When a user taps a key, swipes, or tilts a joystick, the system responds within 40 milliseconds to alter the ongoing video stream.

Unsurprisingly, it is far from polished. Odyssey openly describes the current software as feeling like exploring a glitchy dream: raw, unstable, but undeniably new. Visuals can blur, textures can warp, and environments might suddenly shift without warning.

To keep things from breaking entirely, this early release is a narrow distribution model trained only on specific real-world footage. The team says a next-generation model is already in the works, promising sharper fidelity, better physics, and more coherent interactions.

Big backers and expensive compute

Odyssey has raised $27 million from investors including EQT Ventures, GV, and Air Street Capital. More intriguingly, Pixar co-founder and former Disney Animation president Ed Catmull sits on the company's board.

Keeping this engine running is not cheap. Odyssey disclosed that running the AI tool currently costs between $1 and $2 per user-hour. The startup insists this price tag is falling rapidly thanks to model optimization and hardware tailwinds, but for now, it remains a heavy infrastructure burden.

The startup is also facing massive competition. World models have become the hottest battleground in AI, with heavyweight rivals like Google DeepMind, Microsoft, and Fei-Fei Li’s World Labs all racing to build responsive, physically grounded digital environments.

Why it matters

Instead of assembling 3D assets, polygon meshes, and physics engines, Odyssey is betting that generative neural networks can bypass the traditional graphics pipeline altogether. If a model can simply hallucinate coherent real-time worlds from controller inputs, the boundary between video and video games starts to dissolve.

Right now, we think interactive video is mostly an intriguing tech demo rather than an everyday creative tool. Average users will find it charmingly bizarre, but nobody is binge-watching or playing an unstable hallucination for hours just yet.

The real test will be economics and coherence. Burning $1 to $2 every hour per user makes mass consumer distribution impossible under standard subscription models. And keeping a generative world consistent over extended sessions remains an unsolved hurdle across the industry.

Still, Odyssey made a clever move by shipping an unpolished preview directly to the public. By framing the glitches as an experimental aesthetic rather than a failed game engine, they have planted a very visible flag in a field that tech giants are rushing to dominate.