Runway's CEO argues AI video was the prequel to world models
Cristóbal Valenzuela makes the case that text-to-video is a transitional capability, and that simulating a persistent world with consistent physics is the actual target.
If this is right, the craft shifts from prompting a shot to directing inside a scene — blocking, camera, and continuity become available again rather than being things you fight the model for. That is a return to something much closer to conventional filmmaking, which is good news for people who already know how to do it.
When the chief executive of one of the best-known AI video companies describes AI video as a prequel, it is a statement about where the company intends to spend its money. Cristóbal Valenzuela’s appearance on TechCrunch’s Equity podcast frames text-to-video as a step toward something else, and that framing has consequences for anyone who has started building work around today’s clip generators.
TechCrunch’s episode notes outline the territory rather than the full argument. Runway, they note, has raised close to $860 million at a $5.3 billion valuation and is pushing into general world models with applications in gaming, robotics and possibly something closer to general intelligence. The conversation covers why Valenzuela thinks the real constraint on filmmaking has never been technology, how Runway approaches world models differently from Google and other labs, what he calls “nonlinear media,” why real-time video generation opens use cases well beyond content creation, and his view that AI companions are not inherently dystopian. The episode page does not spell out his answers in detail, so the substance here rests on those stated themes and on what Runway has shipped.
What Runway has shipped gives the idea a concrete shape. GWM-1 in December 2025 worked by frame-by-frame prediction with an internal sense of physics, and GWM Worlds 2 in September streamed continuous 720p video at 24fps, steered by text actions and camera movement, with Runway listing its own limits: detail degrading on fast camera turns and imperfect long-term memory. That is the practical gap between a clip and a world. A clip only has to look right for its duration; a world has to remember that a door was open, that a character was holding a cup, and where the sun was when the camera comes back around.
For filmmakers, that distinction maps directly onto craft. If a scene persists, blocking and coverage become possible again: you can place a camera, move it, cut to a reverse and expect the space to hold. That resembles the virtual production and previs workflows that already run on game engines more than it resembles prompting, and it would favour people who already think in terms of eyelines, continuity and lensing. The nonlinear media idea points the other way, toward work that is explored rather than watched, which is closer to games than to film and would need different deliverables, review processes and rights structures.
The caveats are real. Persistence is exactly what current systems struggle with, as the NVIDIA and Princeton research on addressable memory in August showed when it found that world models lose track of earlier scenes once a rollout runs past their training length. The term itself is unsettled, which is why a July paper attempted a definition and roadmap. And the competition is heavy: Google’s Genie 4 in October doubled scene consistency to two minutes, and AMD agreed in September to buy World Labs for $8.2 billion. Whether a world model can deliver a finished frame, rather than a navigable sketch, is the question to watch.