Okay, we're joined by Cristóbal Valenzuela, co-founder and co-CEO of Runway. Thank you very much for your time.
Yeah, of course. Thank you for having me here.
Yeah, you just got here or
Couple days ago. Okay. How long are you here for this time?
Uh just like over a week, a little bit less than a week.
Okay. And how do you come to Japan quite often or?
Uh it's my second time. Uh the first time I came here was almost two months ago. Um so hopefully I'll come more often now. Um at least a couple times a year. Mhm.
Okay. Just to give our audience a quick introduction to yourself and the company. Um here's the AI generated clip of a lamp that can walk and interact with people. Uh this has been published a few weeks ago, I guess. Yeah. And this is made by Runway's arts.
So yeah, this is really photo realistic and I can't believe this is AI generated. But what does Runway do essentially?
Yes. Um, so what we do in short is we help companies and we help creatives tell stories. And the way we do that is we develop products and technologies that allows you to generate and create any sort of content that you want. And so what you're seeing on screen right now is kind of a demo of that. This is a story that we made internally. It was done by one person in the company and it's really to showcase a little bit of the diverse and complex capabilities of what AI is able to achieve these days. And this is just the beginning.
I mean really a touching story but it's also the worst it will ever be.
How does people make this like just short prompt and this was made or chat with the models or what?
So there are different ways of making content in Runway, depends really on like what your intent is. Um for to get something like this, you know, where there's a lot of creative and creativity and human ingenuity behind it, you have to know really what you want to tell and then what you use Runway for is to prompt and use the models and use our systems to get you there. And there's no one single answer. There's no one single prompt you should use. Everyone uses differently. Everyone has different techniques.
It's really a general purpose technology.
And so the creative who made this has a particular way of working but someone else might have another one. And so that's really kind of the idea is that you can explore and have diversity of use cases and there's no one single way of using it.
You said general purpose. So it's not just for filmmakers or videographers. And so who uses this?
Yeah. So AI and Runway I would say just generally are one of the few really transformative general purpose technologies. And so what do we mean by general purpose? You know the internet is general purpose technology. The internet has so many wide set of applications and it can be used for so many different things and the really the AI revolution that we're in right now is in part due to AI being able to generalize and have applications in so many diversity of fields. And so when you look at Runway and the kind of work that we do, a lot of people think that it's great for filmmaking and for content for advertising which is true but at the same time the tools are way more generalized and way more kind of broad than just that set of creative use cases. And so today if you look at a customer base, if you look inside the data, we have users in design. We have users in brands, in consumer goods, in financial services, in gaming, in architectural kind of companies, in education, in sales. It's basically you start realizing that really AI as a general purpose technology is allowing you to tell stories. And well, who tells stories? Of course, filmmakers and Hollywood tell stories, but really every company out there is trying to tell a story. And not every company, I think every person might want to tell a story. And so for us, it's really understanding how do we create tools that will allow even a larger set of people in the world to tell a story that makes sense to them.
Mhm. So not just like creators but also like marketers or sales or I don't know the general people?
Well for me those are also creators like I think creativity for me is a state of mind. It's a way of looking at the world. And sometimes when we think about creators or creativity we tend to think of artists you know and a small segment of the world that works in content and film. But I think you can be creative. If you're a salesperson, you can be incredibly creative about how you approach your leads. Or if you're designing like a shoe or you're designing a piece of clothes, you can be extremely creative about how you think about the choices that you make. And so our purpose is not only to serve, you know, the filmic and the kind of creative community that we might traditionally think of as creatives, but really enable that creativity that's latent in many more people. And for example, a good example of this is we run a film festival every year across the world and we're actually bringing it to Japan this year. And if you look at many of the winners of these, are people who make videos or make films and have won the festival. There are sometimes, they're not coming from a filmmaking background. They're not filmmakers themselves, they're musicians or there are people who work in brands and now they finally understand that they can use these tools to tell and make things they never thought they could. And that's really like the most interesting thing I would say in AI is that there's latent possibilities that are there dormant until you have the great technology, a good set of technology, you can find ways of expressing that latent creativity that's always been there.
Mhm. So like that means AI enables like literally everyone to like make videos easily.
Correct. Yeah. I mean that's the role of a general purpose technology but not only videos. I think videos it's kind of one medium of expression but there's many other things you can do with Runway. It's more of a multimedia kind of platform. But you're right. I think that's the purpose. Like you can think of writing. Writing was many hundreds of years ago, a very small set amount of people knew how to write. And as technology got better, as the printing press got better, more people managed to read and write. I know writing is like, well, everyone can write, right? Everyone has benefited from the fact that you can write great stories and share them really quickly. That hasn't like realized in the visual domain because of course making AAA movies is not as cheap as writing a one page document, right? But our idea and our vision is that at some point it will, at some point making a AAA movie will be as cheap as just writing a document on your computer.
Okay. So first about Runway and the evolution of AI video. You founded a company back in 2018.
And that was way before ChatGPT.
Oh yeah, way before that.
So there was no major LLMs and no major AI video models. Could you tell us how you started developing the video models since you founded the company?
Yeah, of course. So, you're right. We started the company in 2018 and the idea of it was, you know, we were fortunate enough to be at a time early on where we started to see firsthand the progress deep learning was making. You know people think AI is almost like a more recent development but the truth is that AI has been a topic of research and engineering interest for the last like 60, 70 years and there's like waves of a lot of interest and then like a lot of less interest and those are called AI winters and the latest revolution of AI is on the heels of a specific technique called deep learning and deep learning really started to work around 2014, 2015. And as founders, we were three founders who were fortunate enough to both be paying attention to those very early developments and also we were interested in that. And so we met at school. We studied at NYU from 2016 to 2018. And a lot of our research at the time was focused on how do you use these deep learning techniques but in creative ways. And that might sound obvious now, but this is in 2018. This is like prehistoric times in AI. This is where AI is moving so fast these days that eight years ago really feel like a hundred years ago. So we really didn't exactly know how we're going to get there. But we knew that these algorithms, these deep neural networks have the potential to effectively kind of do what we were just discussing which is augment people's creativity. And that was actually the tagline that we started the company with was Runway augmenting your creativity with AI. And so between 2018 and I would say 2022 we started just experimenting a lot with different techniques and products and research. And around 2022 was the first time I think we had both convictions and results and some techniques and ideas that were working specifically around image generation. And 2022 was the time also ChatGPT came around and I think ChatGPT and the amount of progress that was happening at the time was just a great catalyzer of a lot of things.
A year after that you released your first video model. Gen-1?
2023 was like the first video model correct. And that was the first time, you know I remember having a conversation with a friend who also runs a big company in the US and this is like late 2022 when we're discussing, you know this is the beginning of LLMs, if the beginning of large image models and a lot of other things and he was asking me, you know when do you think we're going to be able to do video. And I knew that it was coming like in a few more weeks because we're working on it. And I was like well hopefully soon. And he was like no it's going to happen in like five years. And so I think that has been like a good perspective of seeing the evolution of what you can do these days and also is a good reminder of, you know, sometimes we forget how far the industry has come.
I almost think that we are in a time, in a moment where we've normalized magic. If you think about what you can do with AI, what you can use Runway for, what you can use language models for, if I told you that you could do that five years ago, you just wouldn't believe it. Like no one would have believed that we would get to a point where you can write a prompt and get a hyperrealistic video of what you just described. But we can. And that for me is like a magical moment of we've normalized this very unique technology and this very magical kind of set of things that we can do and we're getting used to it which is actually kind of great.
Yeah. So when I think of the other companies that already developed the video models like Google or OpenAI, they're a huge company now but compared to that Runway is a much smaller company and OpenAI and Google have prior LLM development. So how could Runway, a small company like Runway, achieve those development of models?
Well, this is like if you look at the evolution of innovation, why companies win particular markets if you remove AI from the equation and just look at why startups are managing to disrupt large incumbents. I mean Google was a startup at some point and they managed to overcome every single search engine that were much bigger than all of them and the reason is I think, well there are a couple things. One is focus. I think you have to be a very focused company. You have to have one consistent mission and be able to achieve that mission consistently and be willing to go through the pain of doing the things that work and the things that don't work. And I think that's a state of mind and I think when you're a startup and you're a company, size might not matter if you're not focused on something. The second thing I would say is speed. Of course innovation is driven by focus and by care and thinking about problems differently but also in today's world I think if you're slow or not fast enough I think it doesn't really matter. And I think when you put a larger company next to what we do, I think we've managed to always hopefully continue to win because of our speed and focus. And I think that's been a constant flywheel on the way we work. And that's a constant pattern how startups continuously win over large incumbents.
It's the innovator's dilemma. If you can't figure out how to change, you're going to die. Mhm.
We still remember the AI generated video of Will Smith eating spaghetti like it was messed up.
But somebody compared the past video and the current video on X or something and then it works smoothly now and a lot more realistic. So how did the technology itself get to where it is right now?
I mean, if you think about Will eating spaghetti, that was like two years ago. Just two years. That's my point. Like if you look at the progress of computer graphics over the last 40 years, you need to look at what computers were able to do in the 70s and then compare it to what a Pixar movie looks today. And in 40 years, that domain has overcome and evolved a lot. And so we're seeing kind of a similar transition, but in two years. So what a computer was able to do two years ago feels like 40 years in traditional evolution of graphics, you know. And I think the reason it happens so fast and what happens to get to a point where now you're able to create hyperrealistic videos of content or anything you want, I think it's a combination of well like better larger models. I think scaling laws for video models have become kind of real and you want to train better networks to just scale them, more compute, more data. But secondly, there's a lot of new algorithmic improvements and techniques on how you improve them. And maybe to the point of how do you succeed as a smaller company is that you're innovating faster. You're discovering the things that work really fast. And I think compounding that has been the result, if you compound that, that's why I think models have gotten to a point where they actually work and do what they're able to do these days.
How did you secure the compute cost on the road, in the journey?
I mean you have to raise money, you have to work with partners. We've been fortunate enough to find great partners who can help us scale, who believed in that mission very early on. And it's about building great products and with those products hopefully drive more attention, get more customers and just keep growing the business so we can invest most of what we've created into more compute.
So if I look at how much you guys fundraised, that was pretty much how you guys invested in the compute.
Yeah. I mean if you're an AI lab these days or research lab you should expect that most of what you're raising is going to compute. It's how you train the models and how you make sure you continue making progress. So yeah, I think that's not controversial even for any AI lab who wants to create new models and frontier models.
But you're not building data centers by yourself, right?
We're not building data centers by ourselves now. But I don't know, maybe at some point we'll consider something like that.
OpenAI announced Sora first back in 2024 and they launched the app last year and they ended it this year. So it was partly because of compute capacity. So what was your reaction to OpenAI's move like launching their models? Was it a threat for you or?
So look at the beginning I remember like the day they launched Sora it wasn't a very impressive model by the times and the standards, this is as you were saying 2024, I think we had I think one or one of the two only other video models out there and so having first of all a great company like OpenAI come and try to compete in your market I think it's a great sign in a way, it means that this is an interesting market for more companies to start paying attention to. But at the same time I think they deprecated it a couple months ago and I think that for me is a consequence of focus, like you can't attempt to do one hundred different things specifically in this market that's very competitive. Focus on one thing you're really good at and be the best at it. And I think that that's a consequence of lack of focus I would say.
And Google also announced the multimodal model Gemini Omni back in May and they have the creative editing tool Flow. And so they're more interested in the video production sort of side of AI and Chinese companies are also working on video AI models like Zhipu and others. So what does the competition look like to you now?
It means that when you create a category and the largest companies in the world want to compete with you, it means that the category you've created is not only interesting enough, it's really important. And so almost like you want to feel and you want to know that you will have great competitors if it's a great market. And so it's kind of an honor to have ByteDance and Google also trying to get more into this market.
When I talked to one creative director in Japan he said that the launching like products like Flow for Google means they want to train on the data of how actual creators create content. And do you agree with this or the editing tool is more important right now for the creative industry?
One of the things that really matters when you're deploying these tools is the feedback loops that you get when you put this into the hands of people who have never had the chance of making content at this level before. I think that for me is less of the data and more of the patterns that you see emerging from aggregations of users. What things tend to work better, what techniques tend to be different. As a great toolmaker, you should enable those discoveries and these new ways of using the tools even in ways you haven't thought of before. And so what you want to do more than gathering data is enabling more of that discovery and experimentation because in experimentation there's a key in how new things can be built and that's I think what we need more of these days.
Google is also an investor in Runway right but also a competitor. How does the relationship work?
