The demo looks, at first, like a glitchy first-person video game. A wooden cabin in a forest clearing, branches swaying, a path leading nowhere in particular. You press the W key. The screen updates. The trees ahead drift past your viewpoint as though you walked into them. Nothing about that sequence requires a game engine. No textured 3D model exists in memory. Each frame, generated 25 times per second, is being predicted by an AI model whose only inputs are the previous frame, your keystrokes, and the model’s learned sense of what the world should probably do next. The London-based startup Odyssey calls this interactive video. The label may not stick, but the technical claim behind it is harder to dismiss.
What Odyssey shipped, and why it matters
Odyssey released a research preview in May 2025 of its first interactive world model, Odyssey-1, billed as an early version of the Holodeck and accessible free of charge for anyone whose internet connection and patience could tolerate the load. The model generates and streams new video frames every 40 milliseconds, responding to inputs from keyboard, phone, or controller. The company has since released Odyssey-2, an evolution of the same architecture that delivers improved temporal stability and longer coherent generation runs.
The founding team’s pedigree is worth attention. Oliver Cameron and Jeff Hawke, the co-founders, came out of the self-driving vehicle industry, where Cameron had led the autonomy team at Voyage and Hawke had been a senior figure at Wayve. The translation from self-driving research to interactive world models is less arbitrary than it appears. Both fields require AI systems that can predict the immediate future of a visual scene conditioned on the agent’s actions, with low latency and acceptable physical plausibility. The shared technical problem is what attracted Edwin Catmull, the Pixar co-founder, to join Odyssey’s board and invest in the company. The endorsement from the person who arguably defined how computer-generated imagery scaled into cinema carries weight beyond the typical startup advisor announcement.
The architectural problem the team solved
Building an interactive video model is harder than building a generative video model because the system cannot cheat. A clip generator like OpenAI’s Sora or Google’s Veo can look at the entire intended trajectory of a scene before committing to any particular frame, optimizing for global coherence at the cost of latency. An interactive model has no such luxury. Every frame must be generated conditional only on the frames already produced and the user actions already received. The system must be causal and autoregressive in the strict sense, predicting one frame at a time with no knowledge of what the user will do next.
This creates two compounding problems. The first is drift: small errors in each predicted frame accumulate over time, and within minutes the generated world can dissolve into incoherent texture. The second is action grounding: the model has to learn what each user input is supposed to mean, often without explicit labels, by inferring patterns from large datasets of video where someone or something was moving through environments.
Odyssey’s approach to drift involves what the team calls a narrow distribution model, namely pre-training on broad video footage and then fine-tuning on a smaller set of specific environments to constrain the generative space. The trade-off is reduced variety in exchange for stability. The company’s stated next-generation model is designed to relax that constraint, pursuing what Odyssey describes as richer pixel quality, dynamics, and actions. The path is the same one every diffusion-adjacent video research program is now pursuing, and the patterns echo the architectural shifts we documented in our diffusion models 2025 analysis and the AI image generation coverage.
For action grounding, the team has gone unusually deep on data collection. Odyssey designed a 360-degree backpack-mounted camera rig that captures real-world environments specifically for world-model training, on the theory that publicly available video data is insufficient to train production-grade models. The hardware bet is consistent with the company’s broader thesis: better models will come from better data, not just better architectures, and proprietary capture is the leverage point that small labs still hold against the hyperscalers.
Where this could land commercially
Odyssey’s official framing emphasizes film and game production, but the demo applications surfacing from early users suggest a wider set of possibilities. Interactive marketing, where a brand experience adapts to user input rather than running through a fixed script. Virtual tourism, where a user can explore a location’s footage in any direction rather than along a pre-recorded path. Training simulations, where the cost of producing a richly interactive environment drops below the threshold that justifies dedicated game-engine development. Education, where a virtual field trip to a historical site becomes plausible without a production team.
The infrastructure side is less glamorous and more constraining. Running interactive video at 40-millisecond cadence requires significant GPU capacity, and the per-user cost in 2025 was reportedly high enough that Odyssey’s free preview was rate-limited by GPU availability rather than by business model. The economics will only work at scale when inference cost drops by roughly an order of magnitude, a trajectory that the broader generative video field is on but has not yet completed. The competitive set Odyssey faces here includes Decart, Microsoft’s research division, DeepMind, and Fei-Fei Li’s World Labs, all of which are pursuing variations of world models. The category is real. The market positioning is still being decided.
The dynamics surfacing in this space connect to the broader pattern documented in our Tencent Hunyuan Video Foley coverage, where audio generation is closing the realism gap on AI-generated video, and the AI agents analysis, where agentic systems are increasingly expected to operate in simulated visual environments before being deployed in the real world.
A reorientation worth naming
The reframing that interactive video forces is structural. Traditional content production assumes that the asset is produced first and consumed second. A film is filmed and edited, then released. A game is built and tested, then shipped. Interactive video collapses that sequence, because the asset does not exist until the user starts consuming it. The implications for content rights, production economics, and creative authorship are larger than the current consumer-facing demos suggest.
For production studios, the procurement question shifts from “which engine and which assets” to “which model and which fine-tuning corpus.” For platforms, the distribution question shifts from “stream a fixed clip to many viewers” to “generate a unique clip for each viewer in real time.” For regulators, the provenance question shifts from “did this footage capture a real event” to “what trained the model that generated this footage, and is the user being misled about what they are watching.”
These are not problems Odyssey alone can solve. They are problems the category will have to address collectively, and the labs that build the most credible answers will absorb the buyers willing to pay for legitimacy. The patterns surfacing here parallel the procurement realism documented in our enterprise AI governance coverage and the agentic AI report.
The question for builders watching this space
The interactive video category is at an awkward stage. The demos are good enough to be exciting and not good enough to be production-ready. The infrastructure is real but not yet affordable at scale. The legal framework is ambiguous in ways that will eventually resolve, but not on a predictable timeline. For technical leaders evaluating whether to build on this category now or wait, the more useful frame is not “is this ready” but “which capabilities will be ready by the time my product needs them.”
So the question worth putting to any team considering an interactive video integration in 2026: if Odyssey or one of its competitors shipped a production-grade model at one-tenth the current inference cost in 18 months, would your product be positioned to consume it, or would your competitors be the ones absorbing the upside while you are still building around the assumptions of 2024?
