World Models Are the Next Big Leap in AI

AI is graduating from static creations to building simulations that feel alive, reactive, and convincingly real.

At the recent Web Summit, Cristóbal Valenzuela (founder and CEO, Runway), argued that World Models – AI systems trained to understand and predict the physical rules of our universe – represent the next major leap. He described them as a “new kind of camera” for reality.

For developers, this shift represents more than a model upgrade. It opens entirely new classes of applications, moving from simple content generation to real-time, interactive simulation.

World Models let AI understand and predict reality

Valenzuela traces a clear progression in generative AI. First, photography and language models (LLMs) act as tools that represent or abstract reality, with language models capturing patterns in human communication, essentially a human-created abstraction of the world.

Next, video generation models, such as Gen-1 and Gen-2, produce contiguous visual narratives, initially focused on storytelling and entertainment. These models maintain temporal consistency but primarily operate on the visual surface.

Finally, world models go further by training not only on pixels and text but also on the principles of physical reality, 3D geometry, and cause-and-effect. They learn how the world behaves and can predict actions and consequences within a simulated space.

Just as we humans mentally model the trajectory of a falling object after observing gravity, World Models develop their own mental model of the physical world. They train on diverse inputs (video, images, text, and natural observations) to capture reality’s complexity without relying solely on language.

New Verticals and Real-Time Experiences

Shifting to models that simulate entire worlds instead of generating single scenes affects developers in several ways:

Real-Time Interactive Experiences

World Models enable real-time inference, moving Generative AI from the render queue into live applications. Developers can generate non-linear, interactive game worlds on the fly, creating emergent gameplay instead of pre-scripted narratives. Educational content, tours, and product demos can adapt in real-time to a user’s query and context, with the pixel stream generated dynamically rather than pre-recorded.

Revolutionizing Robotics and Data Collection

AI for robotics and autonomous systems often stalls at data acquisition. World Models create synthetic training data, simulating scenarios that would take terabytes of real-world footage to capture. Robots can “watch and learn” in rare edge cases or complex environments without expensive, time-consuming data collection.

New Workflow Paradigms

Content creators will shift from editing to prompting reality. Tools like Runway’s Gen-4 already improve narrative consistency across scenes, paving the way for full World Models. Developers must move from rigid pipelines to flexible, generative workflows, letting models handle physics, lighting, and consistency based on high-level direction.

The main technical hurdle remains compute capacity

Valenzuela believes early forms of World Models are “here today” and expects major improvements in consistency and inference speed over the next 12–18 months.

The main technical hurdle remains compute capacity. Generating fully consistent, interactive worlds in real-time demands immense GPU power, requiring optimized code and efficient resource allocation.

For developers, the challenge goes beyond technology. They must embrace a new paradigm and build applications where pixels emerge from the model’s understanding of physics, not static assets.