For three years the whole industry optimized one thing: models that predict the next word. The frontier just moved. DeepMind, NVIDIA, World Labs, and Yann LeCun's new lab are now racing to build models that predict the next state of the world and what actually happens when you act on an environment, not just what a sentence should say next. LeCun left Meta over this. The money and the talent are following.
The word for it is world model, and it's worth being precise about what one actually is, because the term is already getting stretched over anything that outputs a picture.
What a world model is
A world model does two jobs. It holds a representation of a system, the entities in it and how they connect. And it runs that representation forward to predict what the system does next.
The second job is the one that matters, and it's the one almost nobody ships. Genie 3 generates gorgeous interactive worlds, but there's no real structure underneath the pixels, so it's a renderer, not a simulator. NVIDIA's Cosmos is the real thing for physical AI: it produces physics-accurate video that robots and self-driving stacks train against, so they can practice a tornado or a warehouse collision without one ever happening. Fei-Fei Li's team spent a whole essay this year just trying to stop the term from collapsing into "any model that outputs a scene." The distinction they keep drawing is representation versus simulation. You need both, and simulation is the hard half.
One thing every one of these bets has in common: they're all aimed at physical space. Robots, roads, game worlds. Almost nobody is pointing this at the environment most engineering orgs actually spend their nights fighting.
A context graph is only the first half
Every observability vendor has some version of a context graph now. Services, dependencies, ownership, recent deploys, past incidents, all wired together and queryable. It's genuinely useful, and it's where the market has settled.
But a graph only knows the present and the past. You can ask it what depends on the payments service, or what broke the last three times latency spiked. You cannot ask it what breaks next if the payments service starts degrading right now, because a graph doesn't run. It has no notion of time moving forward. It answers lookups, not "what if."
A world model runs the same data forward. That's the entire difference, and it's a large one.
The industry already proved the second half works
Nobody has aimed this at infra yet, but the physical-AI companies have already shown what forward simulation buys you.
Waymo trains against edge cases generated by a Genie 3 variant instead of waiting to meet them on the road. Figure, Waabi, Agility, and Uber all run Cosmos in production to manufacture the training scenarios that are too rare or too dangerous to collect for real. The bet underneath all of it is the same: simulating the expensive failure beats waiting for it.
SRE has been running a cruder version of that idea for a decade. It's called chaos engineering. Break something on purpose in a controlled way so you learn how it fails before it fails on its own. A world model for infrastructure would get you the same payoff without the blast radius. You wouldn't inject the failure to see the cascade. The model would have already played it out.
Why infra is harder to simulate than roads
Physical world models get physics for free. Gravity and momentum work identically in every scene Cosmos generates, so the model can learn one set of rules and apply them everywhere.
Infrastructure has no equivalent. Every system fails in its own way, the rules drift with every deploy, and the same trigger can cascade three different ways depending on state you can't fully observe. That difficulty is exactly why the graph is where most tools stop.
But the graph was always only half the thing. The frontier labs have already shown the second half is what changes the game for robots and roads, and eventually for the systems running underneath them. I think it's a good time to introduce it in the infrastructure to make the infra more structural. But yes, we would see how it goes.
Till then, Thank you for reading!
Top comments (0)