I went down another rabbit hole. This time, it was world models. I actually was under the wrong impression about what world models were. I used to think that they were models that sort of encapsulated the entire world, but in a sense, I was wrong, partially right but also wrong.
The understanding that I got from going down this rabbit hole was that world models are basically a physical simulation of our world. It's hard to really understand just how important they are if you only read about them. When you dig deeper, you realize that world models are the next step into this whole realm of artificial intelligence.
You see, when a world model gets built, it has all the physical environmental constraints and phenomena that actually happen in real life. You can simulate a thunderstorm, or you can simulate a fire, a forest fire, anything. The biggest use case for world models that we've been able to come up with so far is that they are great data for training physical AI. By physical AI, I mean robots mostly, right? As we get on and with the advent of more physical AI products: it will be more and more clear why we need them, but the reasons are quite obvious. You would think they are useful for simulating stuff. Robots can basically react in sort of a physical, sandboxed environment to these situations, as we can simulate them on the world models.
For example, if there's a fire, and right now I'm talking about a world where there is a very sizable portion of the population that is humanoid robots. Let's say there's a fire, and if these robots are sort of not trained, I would assume that they would try to save themselves. A very big leverage of having trained these guys on situations like these is that now they know that a fire is likely going to cause a loss of human life. I'm assuming that we're living in a world where humans are still the dominant species among everything that lives on Earth. I would say that the robots might be trained to sort of save human lives, which they are originally intended to do. Something like that. I'm not being able to explain it very clearly, but if you want to take a simpler example, maybe we can take the example of Waymo, which is an autonomous vehicle. Waymos are very popular in San Francisco, where these are driverless cars. You just order a Waymo like you would order an Uber, and it just takes you to your destination.
Essentially, Waymos are trained on some sort of a world model that depicts a non-ideal environment where there is friction and where there is also human interaction in terms of cars being driven by humans. How do robots and autonomous systems factor that in? I think this is a better example.
Why do we need more advancements in world models? Let's take the Waymo example again. You might or might not have read in the news that, very recently, during the Fourth of July fireworks, there was a huge traffic jam caused in SF because these Waymos, which are a large fleet there, just didn't know how to react to the excessive crowds that came out on the streets. They just stopped reacting and they just stopped working. These are the types of situations where you still need some sort of research on how to mitigate and simulate adverse situations, unnatural situations, and unperiodic situations in world models (so that we can train all the robotic interactions that we have in real life on them for a more protected future).