Fei-Fei Li Splits 'World Models' Into Three Functions

Putting structure behind a buzzword

Few terms in artificial intelligence have been stretched as far as "world model." The phrase gets attached to everything from cinematic video generators to robotics research, often with little agreement on what it actually means. Fei-Fei Li, the Stanford computer scientist widely credited with helping kick off the modern deep-learning era and now chief executive of the startup World Labs, is trying to impose some order. According to a StartupHub.ai analysis published on June 20, 2026, Li has put forward a taxonomy that sorts any system claiming the label into one of three clearly defined roles.

The framework, drawn from a World Labs essay, breaks the category down as follows:

• Renderers, which generate visual pixels meant for human eyes to look at.
• Simulators, which output geometric and physical information that machines can act on.
• Planners, which close the loop between perceiving a scene and deciding how to act within it.

In Li's framing, a system that only renders, no matter how convincing the output, does not qualify as a true world model. A polished video clip might look photorealistic, but if it has no underlying understanding of geometry, physics, or consequence, it is closer to a moving picture than to a model of the world. The distinction is designed to separate genuine spatial reasoning from impressive visuals that she regards as marketing dressed up as intelligence.

Where Marble fits in

The taxonomy is not purely academic. It also functions as positioning for World Labs' first commercial product, Marble. As StartupHub.ai describes it, Marble straddles the renderer and simulator categories at once. The system can produce the triangle meshes used for high-fidelity visual rendering as well as the collider meshes that a physics engine needs in order to simulate interactions between objects. By generating both forms of output from a single model, Marble can serve viewing and simulation needs together rather than forcing developers to bolt separate tools onto one another.

That dual capability is the practical payoff of Li's definitions. If a world model is supposed to be more than a pretty picture, then a product that delivers both what humans see and what machines can compute against becomes a concrete demonstration of the argument.

A well-capitalized bet on spatial AI

The conceptual case arrives backed by significant funding. StartupHub.ai notes that World Labs has raised roughly $1.23 billion across two rounds, including a $200 million contribution from design-software company Autodesk in a February 2026 raise. That level of investment signals that major players see spatial reasoning as a frontier worth pursuing alongside the language models that have dominated headlines.

Li has long maintained that "spatial intelligence is as critical as, and complementary to, language intelligence," arguing that machines will need to understand three-dimensional space and physical cause-and-effect to operate usefully in the real world. Robotics, gaming, simulation, and design tools all depend on that kind of grounded understanding in ways that text generation alone cannot supply.

Why the definitions matter

By publishing a precise vocabulary, Li is attempting to do something that often shapes emerging fields: define the terms of the debate before competitors do. Whoever sets the standard for what counts as a real world model gains an advantage in how the technology is evaluated and discussed. The taxonomy draws a clear line between World Labs' approach and the many rivals racing to claim the same territory, and it gives customers and researchers a sharper lens for judging which systems truly reason about the world and which merely depict it.