Runway’s CEO on why AI video is just the warm-up act for world models

AI-generated video has gone from a party trick to a legitimate creative tool in what feels like a blink. And Runway, the New York-based startup that’s been at the center of this shift, isn’t satisfied with just making cool clips. They want to build something that actually understands how the world works.

That’s the gist of what CEO Cristóbal Valenzuela has been hinting at in recent interviews. He sees AI video as a “prequel” — a necessary but ultimately limited phase before we get to what he calls “world models.” Not just generating pixels that look like a cat walking, but actually modeling the physics, the causality, the spatial reasoning that makes a cat walking make sense.

Let’s be real: the current state of AI video is impressive, but it’s also brittle. You’ve seen those clips where a person walks through a door and their arm bends like a pretzel, or the lighting shifts randomly between frames. That’s because these models don’t understand what they’re generating. They’re pattern-matchers on steroids, not simulators.

Runway has raised close to $860 million at a $5.3 billion valuation, which puts it in the same financial stratosphere as Google and OpenAI’s video efforts. That kind of money buys you a lot of GPUs and talent, but it doesn’t automatically buy you a world model. The leap from “generates plausible frames” to “understands physics” is more like a chasm.

Valenzuela’s argument is that video generation is the obvious training ground. You need massive amounts of visual data to learn how objects move, how light behaves, how cause and effect play out over time. Video gives you that. But the end goal isn’t better videos — it’s a model that can reason about the world in a way that’s useful for robotics, simulation, planning, maybe even scientific discovery.

I find this framing refreshingly honest. A lot of companies in the AI video space are pretending that the current generation of tools is the destination. Runway is basically saying “this is training wheels.” It’s a bold stance when your entire business is currently built on those training wheels.

The skeptics will point out that we’ve heard this kind of talk before. World models have been a research topic for years, with limited real-world success. DeepMind’s Dreamer and other attempts have shown promise in controlled environments, but nothing that generalizes to the messy, unpredictable real world. Runway’s advantage might be their access to production-quality video data and the practical experience of deploying models at scale.

But here’s the thing: scale alone won’t solve the fundamental problem. Current video models don’t have a concept of object permanence or gravity. They just learn statistical correlations between pixels. A world model needs something more — maybe explicit physics priors, maybe a different architecture entirely. Runway hasn’t revealed how they plan to bridge that gap, and I suspect they don’t fully know yet.

Still, I’d rather see a company aim for something genuinely hard than iterate on incremental improvements to prompt-driven video generation. The latter is a solved problem at this point — every major lab has a model that can produce impressive short clips. The real value will come from models that can simulate, plan, and reason.

Runway’s timeline for this is unclear, and they’re not committing to anything concrete. But the direction is interesting. If they succeed, we’re looking at a fundamentally different kind of AI — one that doesn’t just generate content, but understands the rules of the game. If they fail, they’ll probably still have a solid video generation business to fall back on.

Either way, it’s a bet worth watching. The next few years will tell us whether world models are the next big thing or just another overhyped research direction that never quite delivers.