Real-Time World Models, Explained by a Founder Building Them

you're going to end up writing code, not prompts anymore. We will see shows branch in real time per session, per user, and we will see personalization to a level that was not possible before. In three years, we will generate applications on the fly, pixel by pixel. How have you changed?

Hi, I'm Josh Liss. I lead media and entertainment here at Nebius. Thank you for joining us with this edition of the Nebius for Startups podcast. I am joined by Alberto Tutti. Alberto is the CEO and founder at Reactor. Reactor is the developer platform for world models and real-time video generation. Define a world model for me.

Yes, absolutely. So as you know, there's a lot of definitions of world models, first of all, out there these days. The way that Reactor defines it is a video model that you're probably familiar with that generates frames like a video. However, it is a video model that runs in real time, which means that it generates say something like 24 frames per second or 30 frames per second, but in real time, live, and also which is interactive, which means that you can interact with the model as it is generating the video. And so that is markedly different from, let's call it traditional video generation model in the sense that it doesn't generate a file and just hands you a file after it has generated it. It is continuously generating frames looking back at the frames that came before. And so that implies a few things, meaning that, first of all, the model is aware of the things that came before. And also that it is potentially infinite length. So you don't have anymore the five seconds, 10 seconds limit that you see often nowadays video models. And to be technical, just for one second, those are called bidirectional, which all that it means is that the model generates every single frame in the video in both directions in one go. World models in the way that Reactor defines them as infinite length, real-time video models, they are also regressive, like an LLM. Meaning that they look back at what they generated before and make the next frame.

It's retaining that history, as you mentioned.

Correct. And they can generate videos that are much longer as well. And so that's how we define world models. It's got other implications, but in general, that's how we define them.

And we're excited to see what all those implications are as this evolves. To be the developer platform in that industry, leading in that way, what does that mean?

Yeah, so it means that you need to build a developer platform that is markedly different from existing gen AI developer platforms. And really, that stems from thinking logically about the nature of these models. Like I said, they are, first of all, real time. And they're also interactive, second of all. You cannot just give them a prompt and wait for them to do something because that kind of defeats the purpose of them being interactive, right? And so their real usefulness can only be brought forward by enabling interactions and interactivity. Existing platforms in general have been built for file generation. That is passive. It's not interactive. Yeah. So like I said in the example before of most of nowadays most famous video