For those who are less familiar, give us some quick background. What is Runway, and what do you do?

Runway is a company we've been working on for almost seven years now. We work on frontier research on AI and media, finding the direction of both science and art. We train very large models. We just released Gen-4, which is our latest and most impactful model ever. It allows you to create short-form content for entertainment, media, or fun. It's a video model that has consistency built in, so you can create all these interesting and unique worlds. So think about Runway as a research company where half of it is an art company as well, and we try to find the spaces between those two things.

Big news this week was around Gen-4. Give us more color around the release itself. What are the biggest differences between Gen-3 and Gen-4, both in its abilities and in the underlying technology?

Gen-4 is our fourth generation of models for video synthesis. It's the best model we've ever put out to create videos with nothing more than words or images. The first release, Gen-1, which was the first model to do video generation, was three years ago or so. Since then, we've been consistently improving the model every couple of months, with three major releases after the first one.

This is the first time ever, though, that we've achieved what we call world consistency. World consistency is a huge deal in video generation. It means you can create consistent worlds, with consistent environments, objects, and locations, all across the generation. That becomes really important because AI video until now was limited to short clips, and those clips had no relationship whatsoever to each other. So if you wanted to create long-form content, and by long-form I mean more than just 10 seconds, which is standard for a video model, it was really hard, because nothing made sense cohesively and narratively. Now it does.

The way we wanted to highlight it wasn't by showing the outputs of the model in their raw format. We wanted to show how powerful they are for creating stories and short films. So the leading aspect of the release was a couple of short films that we made. There was an animated one, and a more cinematic, dramatic film as well.

I had a chance to watch The Lonely Little Flame.

I hope you enjoyed it. It's a four-minute clip.

It felt like a Pixar short, almost.

There we go. That's all generated. And the really interesting point is that if you forget it's generated, that's exactly the point. That's the best benchmark you're going to aim for. I want you to fall for the story. I want you to fall for the emotion and what it's making you feel, not how it was made. Before, it was too much about how it was made. It was too much about the model. It didn't work, and now it works. So I'm glad it felt like that, and I'm glad you enjoyed it, because that was pretty much the point.

It felt really special. You mentioned world consistency for the first time. Was that a Runway-specific innovation, something you and the team had been working on and feel was a breakthrough? Or did something happen technically in the broader AI world to unlock it? How much of each led to this moment?

Definitely. We've been working on video generation for quite some years, and there's a lot of practice we've done along the way that has paved the way for many other things to be built on top, and I think that's great. But a lot of what we do is also try to find the uniqueness of film, the grammars and the languages that we need to build into the models from the ground up. A lot of what we do you probably won't find anywhere else. The reason is that I haven't yet found any other place where you have someone with 25 years of experience making films sitting right next to the best peripheral and the best researchers in diffusion models and transformers. That's literally what happens at Runway.

The cross-pollination and the ideation that come from those two people sitting at the same table, thinking about how to make video models better at prompting with the cinematic language filmmakers are used to: well, we should start with this. And then you start from the ground up, reimagining the architecture of the model itself. That's sometimes our unique product experience. The way you use it sits behind the product itself. You don't notice it, but from the ground up it was thought to always be used as a tool for filmmaking. There are a few breakthroughs in the way we train the model and the way we prepare the development itself, and a lot has to do with that combination of those two worlds sitting together.

If you think about something like The Lonely Little Flame, which I would encourage everyone listening to go watch, what does it take to create that today from the user's point of view? How long does it take, and how many people are involved?

There's one thing that's very hard about making something like The Lonely Little Flame, or many of the other films we make, and it's something we can't really help you with. It's the one part where we can't step in, which is the story. The hardest thing about making The Lonely Little Flame is the script: what it is that you actually want to tell. Everything else is secondary.

What I mean by that is that the film itself took around four days to make, beginning to end, so just a couple of days. Then we improved some shots, because we were actually training the model as the film was being developed. We made it a few times because the model just kept getting better. But it was done by one person. That's an interesting thing we haven't told. Just one person.

It's a story, actually, that the director who made it had wanted to make for 10 years or so. If you haven't watched it, definitely go watch it. Before yesterday, I guess, that movie, a four-minute short film, would have cost hundreds of thousands of dollars to make, maybe even more. A couple of years ago, that was a couple of million dollars to make. And now it's one person sitting at their computer in upstate New York, working a couple of days on this idea they had 10 years ago, using Runway, and it works. So what it takes is nothing more than a good idea, and Runway can help you with the rest.

That's incredible. You and the team have been working on Runway for a long time. It was founded in 2018, long before the seminal ChatGPT moment, and just a year after the transformer architecture was invented. But it seems like AI was always part of your vision in some way, shape, or form. When you started Runway, did you have any idea yet of the technological inflection point that had happened or was happening? Or was it right time, right place all around?

Well, the company's name is actually Runway ML. That's what we used to call AI back in the day.

Back when it was called ML.

That's right. Back then it was called ML. The idea of the company, at least the thesis I had at the time, and I think my co-founders were in the same boat, was about neural networks as we know them. What we were seeing at the time, this is 2016: if you do the exercise and look back at the images you were able to generate in 2016, they're very bad. They're very low-res. There's nothing more than, if you squint your eyes, you may see something.

You were doing research at NYU at the time?

That's correct. I was in the art school, working in between computer science and art. And this idea of neural networks generating art, because for me this is about art: these are pixels, imagery you can hopefully use to tell something interesting. What I was seeing at the time made me realize that this actually feels like a new medium. If you can type words and generate images, if you can conjure and synthesize pixels, and you just extrapolate that a couple of years, it might be that we could get to a point where you can make entire feature films. Ten years ago, that felt like, sure, whatever. But slowly over time, we built on that connection.

If I go back to the first demo we ever made, the first deck we ever used to raise money, the ethos of the company was the way we thought about it: we were calling it synthetic media before it joined AI. That's how we referred to the idea that you can generate content with AI. It was synthetic media. So we thought, if this is a new medium, then Runway should be the camera you create this new medium with, and we're going to be the engine for that synthetic media world. Over time, of course, things have evolved. There are way more interesting things happening, and breakthroughs have happened in research. But the company's ethos and DNA are pretty much the same. Nothing has changed at the core of what we do. Of course the models have changed, and that's great.

It's a testament to you and the team that you've had a vision and stayed very specific on it for a long time. Looking at the next three to 10 years, is it fair to say Runway as a company has committed to storytelling, media, and entertainment for the long haul? Or do you think there are other verticals you'll look to expand into over time?

I've been very public about it. I think of Runway as a media company. We're building the foundations of an engine for creating any story you want. And the moment you do that, you have basically infinite stories to consume and watch, to change yourself with and adore, and to relate to. The best way to think about Runway is that we're building both the research itself and the stories on top. The three films, and I think we have more coming up now, were made to emphasize that. We have an entire team of screenwriters, animators, and filmmakers working inside the company, making things, making shows, making episodic content.

The realization is that we're building this infinite dream machine that can do whatever you want. Therefore, you could also build the stories on top, and you have a superpower because you're building the engine itself. So for us, the core assumption has been that the best way to think about Runway is as a media company more than a tech company.

Can you give us a flavor of where we'll be 12 months from now, at the pace your team is executing? What will a user of Runway be able to do?

Think about it like this. Let's say you watch 10 videos per week minimum. Maybe you watch YouTube videos, there are a couple of things you like, maybe one or two episodes of whatever new show is on TV. I think by the end of the year, you're going to get to a point where at least a third of those, if not more, are completely generated, from episodic content to long-form content to nonfiction.

You think we'll be there by the end of the year?

Yeah, absolutely. With Gen-4 right now, we're just showing it, but I can show you things that were made in two days that look like a studio with hundreds of millions of dollars made them. And that's now, as of literally 24 hours ago, accessible to anyone.

What's the goal from Runway's point of view when it comes to individual adoption, me or my wife going on Runway and creating an interesting short film or something for our children, versus big-time Hollywood media production? Or is it both?

I think both. To be honest, you can own a camera without having to be James Cameron. They're both basically the same device. They're both capturing light and recording it, either on a chemical substance or in a digital format. You both have access to a camera. You're using it to take photos of your babies and a short video you share with your wife, and James Cameron is making amazing, award-winning shows and movies.

If you think of the substrate of AI as a camera, then you can serve pretty much everyone who has a need for storytelling, starting from the casual ones. We have a lot of users who use Runway not for any form of professional storytelling, or even to share with anyone else. They use it for personal creative practice. It's similar to going to the gym. You go to the gym and you exercise. It's for you, it's for your body, and you feel good about it. You're not trying to become an athlete or an Olympian, but you feel good.

Think about that, but for your creative mind. You're exercising your brain. You had this crazy idea: let me visualize it. You do it, fine, you move on, and maybe you do it a few times a day, and it's great. That experience wasn't possible before, because the ability to create something out of thin air wasn't around until just a couple of months ago. So suddenly you can do it, and you have a bunch of people who have fallen into a creative state of mind they never thought they could. I love that. We're not going to judge those by the quality of the stories themselves, because the audience is an audience of one. It's just you watching it, you like it, and you move on.

That's also interesting to me because for a long time, most of tech has tended to be about what you're replacing, or which market you're going to step into. I think the most interesting products are the ones that create their own markets, their own new experiences, and their own new users. And I think part of it is doing this already.