Over the past year, “world model” has become one of the most overused terms in the AI industry. A system that generates continuous video is called a world model; a system that lets a player take a few steps within the frame is also called a world model. Robotics, video generation, 3D, and embodied intelligence—completely different research paths—are all slapping this label onto themselves. The term sounds increasingly grand, yet its meaning grows ever more vague. World Labs previously did something that slightly offended their peers: they established a clear standard for what constitutes a “world model.” Simply generating what you’ll see next isn’t enough; a system must at least know where objects are, how they relate to each other, and how their states change, before it even begins to approximate simulating a world. This is precisely why I wrote this article. The next major milestone for world models may not lie in how realistic the generated visuals are, but in whether the model possesses an internal world that a machine can read, measure, and manipulate. Atlas is far from solving this problem—but at least it has framed it correctly. Many systems today labeled as world models may, in hindsight, turn out to be nothing more than more advanced video generators.