Podcast Banner

Podcasts

Paul, Weiss Waking Up With AI

World Models

In this week’s episode of “Paul, Weiss Waking Up With AI,” Katherine Forrest and Anna Gressel explore the emerging field of AI world models and the many applications of simulating a real-world environment, from interactive media experiences to robotic learning and even weather prediction.

Stream here or subscribe on your
preferred podcast app:

Episode Transcript

Katherine Forrest: Welcome, everyone, to another episode of “Paul, Weiss Waking Up With AI.” I’m Katherine Forrest.

Anna Gressel: And I’m Anna Gressel.

Katherine Forrest: And Anna, you know, we’re sitting here in New York only for a very short period of time because you’re about to go away again. But I just want to say that there has been more rain this spring — I’m not, don’t worry. I’m not going to go into another tadpole story. But there has been more rain this spring than you can shake a stick at. And we’re recording this on a day where it’s about to storm all over again.

Anna Gressel: I’m glad that you know that. I feel like I haven’t looked at the forecast today, if I’m being honest, but we’ll see.

Katherine Forrest: Are you in one of those apartments? Still in these Airbnb’s for like, from your fire in your house. So you probably have to like put your neck out the window and like crane to look up at the sky

Anna Gressel: Actually, no, this one has like huge, beautiful windows, but it’s otherwise kind of a quirky apartment. So we’re all adjusting. It’s a period of adjustment.

Katherine Forrest: All right, today, while I am looking out at like this incredible cloudy day with yet more rain, I’m very excited because we’re to be talking about world models. And the world model that I want to talk about actually has a blue sky.

Anna Gressel: Mm, great. Yeah, I mean, this is super exciting as a topic. I know we’ve wanted to cover it for a while. And world models are almost like a new modality for AI. You can kind of think about them that way. We have language models that understand and generate text, audio models that understand and generate speech, video models, of course, that can generate video clips and so on. And world models are meant to understand how the world works and generate data that represents that understanding, and how cool is that?

Katherine Forrest: Okay, so I want to pause on that phrase just for a second, world model, because it’s actually a really super interesting phrase that explains a lot in the words itself of what it is. A world model. So let’s sort of break that apart and bring it down to earth, literally. And one hint at what a world model could be capable of comes from some of the video generation models that we already have like Sora or Veo. These models really are terrific at generating realistic short-form video clips. And to be able to do that successfully, the models have to be trained, and they have to learn certain things about the world, including, for instance, that what goes up comes down, sort of things about like what we would think of as the laws of gravity. And that things that may disappear from view, for instance, if you go around a corner, they’re not out of existence, they’re just out of view. So a world model is a model that is trained to be capable of accurately simulating a world environment. And the worlds that we’re talking about here could be real worlds that are photorealistic real worlds of real places, or they could be virtual worlds that are either photo realistic or cartoonish.