Thinking out loud.
Building in public.
There's a moment you hit when working on AI where things stop feeling impressive. For me, it came from thinking about what it would actually mean for an AI system to operate in a world.
Read PostSimulation is a byproduct of understanding. We're not trying to simulate physics — we're trying to build a system that understands it well enough that simulation falls out naturally.
Read PostWhat we learned in the first month of building on top of Godot, what we threw away, and the three architectural decisions that are already shaping everything downstream.
Read PostEvery major engine treats physics as a runtime layer bolted onto a geometric scene. That worked for games. It won't work for world models. Here's the architectural gap we're trying to close.
Read PostMost spatial AI research is tacitly optimized for robot training — embodied agents navigating physical environments. We think the harder and more interesting problem is building for humans. Here's why.
Read PostMixed reality headsets give you a spatial display. They don't give you a system that reasons about the space you're in. The gap between those two things is where we work.
Read PostThe term gets used everywhere now. We want to be precise about what we mean — and what we don't. A world model isn't a video predictor. It isn't a physics sim. It's something in between, and harder than either.
Read PostEarly architecture notes from our world model engine build — how we're thinking about scene graphs, physics primitives, and the interface between learned and hand-authored components.
Read PostNo noise. Just essays, engine updates, and thinking-in-progress from the STAR Labs team.