“World models recently made our list of 10 Things That Matter in AI Right Now.”
It’s a bit late to the party, but at least the adults are finally talking about the difference between statistical correlation and actual causal understanding. For two years, we’ve been told that scaling laws would eventually solve everything—that if you just shove enough GPUs at a transformer, it will eventually “understand” that a glass of water shatters when it hits the floor. We now know that’s not how it works. Predicting the next token is a parlor trick; predicting the next state of a physical environment is a different beast entirely.
The industry hit a wall with pure LLMs. We’ve reached the point of diminishing returns where adding more tokens doesn’t necessarily add more “common sense.” A model can write a PhD thesis on fluid dynamics but then suggest you glue a piece of toast to your chest to keep warm. The pivot toward world models is an admission that linguistic competence is not the same as environmental competence.
The goal here is to move from a system that knows what words usually follow other words to a system that maintains an internal representation of how the world actually functions. If a model can simulate the physical consequences of an action before it takes it, we stop talking about “chatbots” and start talking about actual agents. Do we actually want a model that understands gravity, or do we just want one that can fake it well enough to fool a VC?
This is where the marketing gets fuzzy. When people see a Sora clip, they think they’re seeing a world model. They aren’t. They’re seeing a very expensive interpolation of pixels. There is a massive gap between a model that can render a realistic video of a cake being eaten and a model that understands the cake is disappearing because of the act of eating.
It’s like the difference between a chef who can plate a dish to look like a Michelin star meal and a chef who actually knows how heat affects a protein. One is about the aesthetic output; the other is about the underlying process. According to the MIT Tech Review piece, the focus is shifting toward whether AI can actually learn to “understand” these dynamics. Until we can prove the model is using a latent physics engine rather than just recalling a similar video from its training set, it’s just a fancy movie studio.
The real-world friction here is the compute. Training a world model requires vastly more data and compute than a text model because the “tokens” are now high-dimensional spatial-temporal frames. We’re talking about H100 clusters that cost more than some small countries’ GDP (which is usually just a euphemism for ‘we spent a billion dollars on H100s’).
For the average developer, this is a problem. You can’t run a true world model on a 3090. If the future of AI is locked behind a wall of compute that only three companies on earth can afford, the “open” part of the ecosystem is dead. We’re moving toward a world where the “intelligence” is essentially a utility provided by a few landlords.
Here is my take: world models are the only way to actually solve the hallucination problem, but we are currently trying to build the roof before the foundation. If a model has a grounded internal map of reality, it can’t “hallucinate” that a ball falls upward because that would violate the internal constraints of its world model. It provides a sanity check that text-only models completely lack.
However, I suspect we are currently trading one set of illusions for another. We’ll move from “textual hallucinations” to “physical hallucinations” where the AI confidently tells a robot to walk through a wall because its internal world model has a glitch in its spatial mapping.
The industry is currently trading one set of illusions for another.
By Q4 2025, we will see the first open-weights world model that can predict a physics-based outcome in a simulated environment with 95% accuracy without having seen the specific scenario in its training set. If that doesn’t happen, then “world models” were just a fancy way to rebrand video generation.
