AMD Bought a World Model Company

AMD recently acquired World Labs, Fei-Fei Li’s world model startup. When a major chip company spends this kind of money on a two-year-old AI lab, it tells you something about where the industry thinks the next decade is going.

This feels like 2022 again, when everyone was poking at ChatGPT and not quite grasping what was coming. World models are in that same early, pre-obvious phase. The difference is the big money is already arriving.

World models are AI systems that learn the rules of an environment purely by interacting with it; no hard-coded physics, no explicit instructions.

Think of it like a chess engine, but generalized. A chess engine doesn’t know what chess “means”; it just learns, through millions of games, which moves lead to winning.

World models work the same way, except the game is reality itself. You define a goal, give it an environment to explore, and it figures out the rest.

As explained in this Y Combinator explainer, the mechanics generalize far beyond board games. (This is a must-watch!)

Take a rocket booster you want to land. Rather than writing out the complex kinematics and calculating the descent manually, you create a digital twin of the booster and the physical conditions, and then get the world model, through iterations, to learn how to adjust each parameter to land the booster.

This is a major shift in our limiting factor.

Traditionally, solving a problem like rocket landing required encoding every relevant physical law, every variable, every edge case; by hand, by experts, over years…

World models sidestep that entirely. No need to understand the problem well enough to describe it. Just need a simulate and wait. The ceiling is compute (power).

This also opens the door to something more interesting: agentic world models. The idea that a system could chain together perception, imagination, and action to pursue goals autonomously. And boom! Primitive terminators are built!

To wrap, here’s a video of an animatronic crab at Disney doing the SCUBA dance. It’s a decent metaphor for what world models would make possible.

We feed an example of a human’s movements, and the model learns to reproduce those patterns the best it can with its action possibilities. No programmer needed. Just convergence to the closest approximation without human-in-the-loop.