Well, it seems there is no consensus on what exactly a World Model is, and if it had already been figured out, then I would be out of a job—or, in this case, out of a PhD.
The way I see it, or at least the way I can describe it, is that a World Model mimics our way of predicting the future. Not the future in the sense of predicting the next lottery numbers, but the future of the simple actions we take: imagining what to write before we do it, where a ball will land if we throw it, or what a painting will look like if we apply a certain brushstroke, and so on and so forth.
Well, maybe predict is not the right word. Maybe imagination is.
As you can see, the future can be quite easy, at least for us, since we take it for granted. But for a machine learning model, the simplest things are actually quite hard...
From a technical perspective, for those more acquainted with the topic, the future—as some may say—is simply the next state, given the current one and the action taken. What I take from this is that working on World Models is a bit like trying to teach a machine to imagine, just as NLP is teaching a machine to read and write, or Computer Vision is teaching a machine to see.
Now... why is this useful, you may ask? Well, imagine if you couldn't imagine. See? Quite hard. You probably wouldn't do much of what you do now, and the same goes for a machine. While language seems to work, it is merely a trick, a self-steering technique that does not work when you move away from language and into the real world—that is, Robotics.