Imagine trying to learn how to bake a soufflé by reading ten thousand cookbooks, but never once entering a kitchen. You can describe the chemical reaction of egg whites and air with absolute precision. You can recite the exact temperature of the oven in Celsius and Fahrenheit. But the moment you actually put a dish in the oven, you have no idea why it collapsed. You have the data, but you don’t have the world.
This is exactly where the AI industry is currently stuck. For years, we’ve been impressed by LLMs that can mimic the structure of human thought, but they are essentially just very fast librarians. They know where the information is, but they don’t know what the information actually means in a physical sense. The current obsession with “world models” is an attempt to fix this. Instead of just predicting the next token in a sentence, the goal is to create systems that understand the causal laws of reality—gravity, collision, and time.
As noted in the latest The Download from MIT Tech Review, there is a growing push to move beyond static text. The idea is that if a model can predict the next frame in a video or the next state of a physical system, it isn’t just doing math—it’s building a mental map of the environment. (I suspect this is mostly a way to justify the astronomical compute costs of video training). The logic is simple: if an AI knows that a glass of water breaks when it hits the floor, it doesn’t need to be told that in a text prompt. It just “knows” because it has modeled the world.
It’s a seductive argument. We want to believe we are close to an AI that can actually reason about the physical world rather than one that just guesses which word comes next based on a probability distribution. But there is a difference between simulating a result and understanding a cause.
Here is the problem: most of these “world models” are actually just high-dimensional curve fitters. When a model like Sora or its successors generates a video of a person walking, it isn’t calculating the friction of the shoe against the pavement or the shift in the center of gravity. It is interpolating between millions of existing videos of people walking. It is not a world model; it is a very sophisticated movie generator.
Do we actually believe that watching a thousand videos of a falling apple is the same as understanding gravity? Of course not. A human child understands gravity by dropping a toy once or twice. An AI needs a trillion tokens of video data to mimic the effect. This is the difference between an internal model of reality and a statistical approximation of a recording. We are essentially trying to teach a blind person to describe colors by having them read a million descriptions of a sunset.
The industry is trying to hide this distinction behind a lot of fancy terminology. They want us to believe that scale is the solution—that if we just throw enough H100s at the problem (and the GPU bill is terrifying), the “understanding” will spontaneously emerge. But there is no evidence that spatial reasoning is an emergent property of token prediction, regardless of whether those tokens are words or pixels.
It is just a very expensive mirror.
The real test will come when these models are forced to deal with “out-of-distribution” physics. If you show a model a world where gravity works in reverse, a true world model would be confused but could adapt. A statistical interpolator will simply hallucinate a mess of pixels because it has no training data for “upward-falling apples.”
By Q4, we will see a benchmark specifically designed to break these “world models” by introducing a physics anomaly that does not exist in any training set. When that happens, the facade will crack. We’ll realize that we haven’t built a world model at all, but rather a very convincing puppet show. Until we move away from pure prediction and toward actual causal inference, we are just building a faster way to be wrong about how the world works.
