World Models: From Pixels to Physics w/ SpAItial CEO Matthias Niessner

World models have become one of the most hyped frontiers in AI — and for good reason. The technology is reaching a tipping point, it’s attracting some of the world’s best AI talent, and many believe world models could be the unlock for both AGI and general-purpose robotics.

But ask five people what a “world model” actually is, and you’ll probably get five different answers.

That’s because there isn’t one approach. There’s a whole spectrum of world models being built today, each trying to understand and simulate the physical world in a different way.

To cut through the ambiguity, I sat down with Matthias Niessner, CEO of SpAItial, a London + Munich-based startup building one of the leading world models for physical AI.

In this episode, we cover:

The spectrum of world models — from pixel-first video models, to 3D spatial representations, to systems that learn physics and cause-and-effect in latent space.

Why video models may be hitting a wall — particularly around spatial consistency, long time horizons, and real-time interaction.

How world models are actually trained — including the role of massive video datasets, data curation, multimodal signals, and learned representations.

“Looks right” vs. “behaves right” — and why generating a photorealistic world is very different from generating one that behaves according to real physics.

Why 3D consistency matters — and why a world that changes as you move through it isn’t really a persistent model of the world.

How physics enters the picture — from relatively simple rigid-body dynamics to far harder problems like fluids, deformation, and objects shattering.

The limits of video game data — and why learning from existing physics engines can only get you so close to reality.