Quote · The a16z Show
Why Physical AI Is the Next Frontier | Applied Intuition
Where this was said
World Models: From Physics-Based Simulation to Neural Reality
At 1:04:00 · chapter starts 59:30
Erik Torenberg asks about world models, noting their current vogue. Ludwig immediately flags the definitional chaos — at a recent CVPR conference, the term meant something different to almost everyone in the room. He then lays out the spectrum clearly: at one end, classical deterministic physics simulation that models the world's geometry, sensor physics, and material properties with technical artists creating CGI-quality assets — still the core of Applied Intuition's simulation stack. At the other end, purely neural simulation that generates video feeds directly from a neural network, capable of being reactive (the environment responds to the autonomous agent's actions). In between sits Gaussian-splatting: 3D world representations that maintain geometric consistency as camera viewpoints change, offering high-fidelity rendering of real-world environments. The reactive neural end is the goal but faces the alignment problem: how do you guarantee the generated world actually behaves like the real world? Perfect alignment would essentially mean you've solved the universe.
World models span a spectrum from deterministic physics-based simulation to fully neural video generation. The holy grail is a reactive neural environment that responds accurately to an AI agent's actions — essentially a perfect model of physical reality. That remains impossibly hard, but progress toward it makes training physical AI dramatically cheaper.
Frontier AI labs can build trillion-parameter models that take minutes to respond — that's fine for a chatbot. Physical AI runs on actual clock time with hard millisecond budgets. On-board models must be tiny, fast, safe, and deterministic. That constraint is the moat, and it's very real.
