AI Video’s Wild Year – Runway CEO on What’s Next

Yeah, thank you for having me.

Yeah, so the film festival is a festival we've been putting together since, I think, 2022. And it's basically an open call for filmmakers to submit films that are somehow using AI. We started, and it was a small collection of artists. We had, I don't know, 300 submissions. I think that was the first call. It was a very small set of submissions from people. This year, it's the third time we've done it, we got around 6,000 submissions from people all over the world.

And these are mostly short films. And we sold out the Lincoln Center, where we did our premiere.

Beautiful event, venue here in New York. And then we did a second show in LA. And then after that, we managed to partner with IMAX to do screenings pretty much all over the US to show the winning finalists. And so those are happening, I think, this week and next. And then after that, we're gonna open the videos so anyone can watch them online. A couple reflections, I think—well, it's wild to see the growth, I would say, of starting with this very small venue in Chinatown, trying to get people to come to films and AI films and watch what it meant three years ago. It felt niche and small then.

And now you're selling out the Lincoln Center, which is insane. We had some great artists and filmmakers in there. It all feels like, I don't know, a very interesting tipping point in terms of adoption of AI, but also how excited just people in general are around AI and the idea that you can make stuff that really moves you. I think at some point many people were less focused on how it was made and more on the stories themselves being moving, which I think should be the ultimate goal.

I think we're beyond enthusiasm and just entirely in the adoption phase for many. I think for every major disruptive technology, you will have apprehensions at the beginning, questions. Again, it's totally new. It's something you've never used before. And so if you don't put your hands on it and use it, it's going to be very hard for you to have a fully formed opinion around it. And I think people forming their own opinions has started to happen over the last year, two years, I would say.

I would say today, based on how we work and what I've heard and the people that we work closely with, most studios, if not all of them, have some sort of AI strategy or are thinking through it, which I think is a great reflection of how useful the models have become to many. There are many films out there that are using AI these days. You might never know about it, and I think that's perfectly fine, because you want to focus on the story more than anything else.

And there's also many more, I think, creatives and folks below the line who are just very excited about understanding what this means to them. The way I even sometimes speak about it for the VFX community is, it's very intense work that you do when you're working on a movie, specifically on the last mile. And so there's reviews and edits and changes and notes from pretty much everyone involved. And so if you're the person doing the edit and the composite and the details at the end, you're gonna work a lot.

And you're gonna work, like, 24/7, seven days a week. And in some cases, the changes are so hard that you have to spend too much time on every single frame or modifying them. If I can give you a tool that helps you do that faster and better, you might have a weekend off.

And so that really resonates a lot with people who work in the industry because having a weekend off, sometimes when you're in the final mile of a project, hasn't been feasible before. But now with technology, I guess you're going to get there, and AI will help you, and Runway will help you finally have a weekend so you can relax.

Editing, I would say, professional films involves many different stages and parts. You take existing footage and you modify it, edit it, add stuff to it, color grade it, remove things from it, or sometimes you just generate entire new scenes. Most of the science fiction or superhero movies that you watch are pretty much all generated—not using AI, but using traditional methods of CGI. And so where Runway fits in, it's, I would say, in kind of both of those worlds.

You can take existing footage and modify it and edit it and remove stuff, add things to it. We released a model called Aleph a couple of weeks ago. It's a really good model that allows you to do something that I think was previously not possible with AI models, which is you don't prompt the model with just language. You actually put a video on first and you ask the model to modify that video. And so, for many folks in the industry, it's been kind of a game changer.

And then if you want to generate new, net-new, novel stuff that you've never had before, then you can also use the model for that. And that could be called coverage in film, where you have one shot, but then angles of that same shot from different positions. You can generate kind of coverage that you couldn't do before, or just generate, like, B-roll, for example. There's a show that's coming now on streaming from a major streaming platform that uses Runway to create basically the establishing shots and the B-roll of many of the independent parts of the film.

Yeah, I think that's a fair representation. And I think it also has to do with our philosophy of really what we're trying to achieve and our vision. We always spoke about Runway as a new kind of camera. We always speak about it as a new medium. It's a new medium in the way that cameras were a new medium in the late 1800s. It allowed people and artists and a bunch of people to see the world in completely different ways. And the camera gave birth to photography, it gave birth to filmmaking, and so on.

I think, for me, AI is somehow a new kind of camera. And that camera has, of course, obvious applications in the fields where the camera—the real camera—is still useful, which is like cinema and filmmaking and video-making and ads, which has been the first stepping stone for video models and world models to function. But for us, that's the stepping stone. It's the first function. Cameras, in the same way, were first used mostly for the arts. They were mostly used for theater recording and stage recording, and then, of course, film.

But then cameras have a bunch of other applications beyond that. Cameras are now in self-driving cars, they're in space, they're satellites, right? They're monitoring many things, our organs and bodies, and being used in all sorts of different applications. And I think for our kind of research that we do, we see a similar path where you start with the most obvious kind of use cases, which happen to be around arts and media and film. But then the applications of models can go much deeper and beyond that.

So we have now customers using Runway for game design. We have people using it for architecture, for architectural rendering. We have folks using it for e-commerce. There's applications in robotics that we're now going to spend a bit more time on and announce some work that we've been doing there. There's a bunch of different applications of this idea that you can create moving pixels in hyper-realistic ways. And so for us, it's trying to tackle and make sure that we can solve for many other use cases of this new kind of camera.

Yeah, of course. It's huge. All agencies these days have realized how important it is. I mean, there's no way back the moment you can do something that used to take you weeks in a minute, unless you prefer just suffering and going through the pain and spending way more time and money. For many people, this becomes just a fundamental tool of how you make ads. And I think video, and specifically advertisers, are faster to adopt new technologies because there's less of a tendency to maintain what used to work.

It's too competitive, it's too fast, customers want more. And so if you can help them do more, then yeah, they're using it the most. They will continue to use it even more, I would say.

It depends. I think it depends on how fixated you are with the past. I think some professionals are very obsessed with how things have worked for a long time, which I think happens to be the case if you look back at history. Like, people get too obsessed with their craft, with the things they know. I've been working on this for 20 years.

Chris, this is how we do things here. I'm like, yeah, sure, great. But, like, you can also do it differently. There's no rules. Things can be made differently. And so I think there's definitely the hardcore people who are gonna, like—some of them are still stuck in the way they wanna do things. And to me, it's great. I mean, we still have analog cameras, and some people still go and, like, reveal films in dark rooms, and you can still do it if you want.

Of course, that's a choice. But I think for many, it's the realization that this is a new medium, and it requires you to rethink from the ground up how you worked before. And sometimes you're gonna bring some of the things they're used to to this new world, and they're not gonna work well. So I'll give you an example. In traditional editing, NLEs, or in traditional CGI and graphics software, the way you export is you click export and you kind of sit there for hours sometimes just to wait for the thing to render.

And then you need to hopefully make sure that once you finish, you watch it and there's no mistakes. If not, you're gonna go back, make the edits, click render, and wait a little bit more. AI doesn't work like that. You can technically generate 10,000 videos at the same time. It just works that way. And then you can pick the ones that you think are closer to where you need to go and then keep iterating from there. That function of working in quantities is very hard for people who are very attached to their linear way of working.

And so sometimes I see people who have worked in the industry for 20 years click generate once and stay there and wait for the thing. I'm like, no, you can generate as many as you want. It's the same. And it's hard for them first to understand it, but then once you understand it, you're going to start exploring what this means. And I think more people now are falling within that bucket of understanding that this has to function in some way different to how you've functioned in the past.

It's not that hard. I think the people that are sometimes the best are newer generations, the younger folks. We have programs at NYU, USC, MIT, UCLA, where people and teachers are using Runway for their classes. And those are the folks that, for this, this is very natural for them. This is how they've grown up over the last couple of years, using these tools. And I think a common theme there is that they don't have a preconception of how things are supposed to work.

They're looking at it with fresh eyes. And when you're coming at the field with fresh eyes, you can ask questions that perhaps you're not supposed to be asking, and you can explore things that weren't supposed to be explored because they just weren't possible before. And I think the people who had the most fun, I would say, with these tools, and the people who are starting to use it more professionally within films or advertising or professional use cases, are the folks who just come at it with fresh eyes, with no preconceptions of trying to fit this within their previous way of working, but just figure out exactly what's new for them.

And I think that will continue to be the case. Yeah.

Yeah.

Yeah, it was hard. I think it was hard because we had conviction around something that very few people had conviction around. And I think for us, it was worth trying to see if we were right. But for many, I think it wasn't worth even trying. And I think when you're alone doing something that no one else believes in, you have to just be insane to try to keep doing it for long. Specifically, we had so many people reach out and be like, "This is a waste of time."

All these images look bad. I have emails from some of the best investors in the world or the best researchers in the world telling me I was wasting my time, that generative AI or creating AIs for images and video is not a use case. And I think we're partially maybe obsessed with proving that it worked. And every time we heard no, it's like, okay, another one that we need to hopefully prove wrong at some point. And so it became like fuel.

We wanted to make sure we could prove that we were into something interesting here. And I think a lot of it has to do with just having conviction when something is not really hot to work on. I think when it's obvious, it's too late. When things are obvious to everyone, I think we've understood that it just becomes like a commodity. It's too late for it. And I think that also becomes the case for us now, where we're thinking about stuff that I think people look at us and they're like, "No, that's insane."

And we're like, "Yeah, exactly. That means that we're into something." If what we're saying and what we're speaking and what we're doing feels obvious, it's too late. It's not going to matter. And I think that conviction early on has taught us something around just being persistent and figuring things out even when few people believe in it.

So the vision was pretty much the same as today, which is, I was actually reviewing a deck that we had from 2018. I think generative AI is the way people describe AI these days. We used to call it synthetic media back then. And our thesis was like, look, we're now able to generate these very small patches of images. And so the first product was a bunch of models that would allow you to generate or use AI in creative ways, sometimes creating images, but in these very blurry, inconsistent ways.

And our thesis was like, this is going to scale, and the moment it scales, you're going to be able to generate anything you want. And so we started with what was possible at the time, which was very small. It was using GANs at the time and LSTMs for writing text, a set of experiments of products that you can plug into existing software. And so we had a Photoshop plugin that allowed you to generate images inside Photoshop. Before Figma, there was a software called Sketch.

So we had a bunch of plugins in Sketch. Then we had a plugin in Unity that allowed you to create renders, a bunch of different explorations, I would say, how to use AI within creative workflows.

So Transformers, for the most part early on, were just used for language. I think we were using other approaches for pixels and video, and diffusion was kind of, I would say, a big transition from what people were using at the time, which was mostly GANs. And now people are using combinations of Transformers and diffusion systems. I think it was a validation of sorts, like, we're into something. And it took us a while to prove that we were right in a way. I think it also brought a lot more attention to the field.

A lot of companies started to appear. There was way more competition than before. But I think for us it was just another reminder: try to remain focused and remain obsessed with what you know is true. There was a lot of noise early on in 2022. There were too many things going on. And I think at some point it was easy to get dizzy with stuff going on and companies popping up and everyone offering everything. And I think we were like, yeah, just keep doing the thing we know we're good at and everything else will follow.

And so, yeah, I think it was, and still today, it feels like a very competitive environment, which is, in a way, great.

I think probably at the time we released Gen-2, which is our second video model. I think before that we had a bunch of image models and things that I think were working pretty well. But I think there's always this tension of where do you want to spend your time optimizing that previous generation of models and trying to push the frontier. And I think for us, it was like, pushing the frontier sounds way more interesting. And once the models, the video models, started to get really good, and I think it was probably Gen-2 times, that was 2023, which in AI land feels like 20 years ago.

That's, I think, where more people start just understanding how useful these models can become. And of course, we released Gen-3 Alpha a couple of months after that. And that's, again, another spike in usage, another spike in use cases. And most recently, Gen-4, which is the latest model, also feels like—I think every release of a model comes with a bunch of new exciting use cases and new people using the models, and then new learnings of what you can do with it.

So I think on every model release, there's, like, product-market fit in a way. But it's interesting to make sure, for us at least, that we don't want to stay there for too long. We need to push it again and again and again.

I think we started mostly as a product company. I think the first two years of Runway were mostly just building product and then kind of realizing that what was missing from that product experience that we thought we should give users and ourselves was, like, core research. And we just didn't see anyone doing that kind of research. And so we started building it ourselves. And building a research org from the ground up is way harder than I thought it would be. But I think that happens to be the case with pretty much building a company.

And then I think by now we have—I mean, I'm biased—but I think it's the best research team in video, in world models, in image. It's a small team, a very small team, that has kind of pushed the frontier for a couple of years now. And we're competing with, I would say, the best research labs with much more funding than we are. And we still manage to do very interesting research. And so we've now managed to, I would say, balance both. The product ethos of the company is still there, but now we're a really strong research team.

But for me, those two things have to go hand in hand. If you're just a product team, I think you're going to get leapfrogged by research. If you're just a research team, then it doesn't have any real impact in the world. So being able to do both has been kind of a superpower of ours.

That's sometimes the case, but I would say for the best people, and maybe something we do when we hire, is try to find people who can understand a little bit of both worlds. I think there's definitely a simplification of thinking of researchers just as academics who want to publish. I think many of them are seeing their impact in the real world and want to build products. And there's a lot of great engineers who understand research and want to make sure they can bring that research to their product development.

I think it's just picking the right people. There's definitely a lot of people who are just obsessed with one or the other. But I would say these days, the intersection of folks who can speak both languages is a bit more common. So you just try to find those people.

So in a way, they're the same model, slightly different. What I mean by that is our thesis has always been that models, as they scale, will start generalizing. And what I mean by that is there's many tasks in video that you rely on very specific specialized models to do. So you were asking me about what kind of things filmmakers use Runway for: all these very specific workflows and things you're using other tools for. The way you could solve for that in AI land before, a couple of years ago, was, like, you build a specialized model for green screen, or a specialized model for inpainting, or a specialized model for avatars, and you have these specific specialized models that do specific things.

Our thesis has always been that that's not going to matter. None of those things will matter the moment you have a model that can learn how to do all of those things at once. And so Aleph is our first public approach towards solving that. The model has this thing that we call in-context. And so what it basically means is that you can solve and do things with the model that the model wasn't specifically trained for. And that gives you so many superpowers because then, if you want to do something, you just have to show the model what is the thing you want to do, and the model will learn how to do it.

Very similar to how you approach perhaps similar problems in language these days, where, I don't know if you remember this, but before the very large language models that we have today, you used to have very specialized models for translations. There was a model in Hugging Face that translated from Korean to English, or there was a coding agent model that used very specific things. None of those models really matter right now because you have a much better model that has learned how to generalize around all of those tasks.

And you can just system prompt the model to behave really well in one particular thing. For example, I think that is, for us, kind of the future, I would say, of video and just world media building, is you don't have specialized models. You have one model that then, if you provide the right references or the right system prompts, the model can tackle those things. And so Aleph and Gen-4 are steps and stepping stones towards realizing that. And they're in some way the same model, just with slight changes on top.

I think we're going to continue pushing the frontier of having these models that can generalize and do a bunch of different things. We do sometimes other things where we specialize our models for particular very narrow tasks, like character performance. So we have a thing called Act-Two, which allows you to—by the way, for podcasts, it's great because you can change people and faces and characters as you wish. And in that case, we sometimes need to match the voice of the person giving the speech, and the expressions do matter a lot.

And so you can take the base model and kind of think about it as, like, you take the base model and you do a system prompt very specifically, or fine-tune very specifically for that task. But still, the underlying model is basically the same.

That's the most interesting thing, I would say, with models that can generalize, is that you don't have to customize the product experience to tackle those specific use cases. And I think that speaks a lot more about the overall trajectory and, I would say, direction of where software is going. I think software, for me, over the last two decades, has been about picking verticals. You pick a specific function of something you want to do, and you just go very deep into it. And so you had Adobe building very specific software for creatives, but you take some of that same engineering work and you build Autodesk, and you've had very specific workflows for architects.

And those two things are different. And then you have gaming world and software, and it's very specialized and specific. But then if you look at something like Aleph or Runway, we don't have very specific software to address all of those needs. It's the same UI, the same product experience, yet somehow we have customers in all those verticals. And I think the underlying, I would say, pattern there is it used to be the case that you pick verticals. I think now you pick principles, and the principles allow you to scale much better than any other previous generation of software.

Our principles are the ones I was kind of telling you before, which is generalizable models. Scale really matters. Data quality really matters. If you pick those principles well enough, you will scale in ways that you couldn't scale before with traditional software. And so, yeah, we don't customize or change the underlying product in any way. We do a lot of the work in the model, and that will continue to be the case.

No, you don't. And that's the interesting thing. You ask the user to do it at inference time, similar to, I would say, language models these days, where if you think about it, I'm sure you use any chatbot these days or language models to work. You're using the underlying same model interface that someone in Chile is using to do their high school homework, and someone at MIT is using for biology research. And it's the same underlying interface. I think for video and image models, it's basically the same.

You don't have to have specialized UIs. You have to have the user just give you the right instructions. And if the right instructions are set correctly, then if there's a need for a UI or a slider or something else, I also believe that the models and the products should be able to just generate that as well, which is a completely different paradigm from how we built software before, and I think a much more interesting one for me.

You can now generate video or content or media without having to write a single prompt, which, for many, I think, is just totally new and goes back to kind of what I was saying before. People might need to rethink their mental models around media and videos and images. Aleph works in a way that allows you to input a video and then either select or choose or make a modification with it. And so you get the same video. So you can generate an entirely new sequence, but in this case, it could be the same video just with something added or modified in the way that you want.

And in some cases, just a word. In some cases, you can just annotate on top, which happens to be the way most professionals work. You take a video or a frame and you annotate on top what you want, and you use that to prompt. So you have a video, your annotations that happen to be images. There's no language prompts. There's only words, perhaps, in the reference or in the annotation. And you feed that to the model. The model looks at both things at the same time.

Understands that this is an annotation of this thing and then makes the changes.

Literally, it doesn't have to be words. Again, if you look at how professionals are working these days, it's pretty much like that. You start annotating on top, and then the annotations get sent to someone. They review it, they interpret them, and they make the changes. That process can now be automated, having a model that does it for you. That, for me, is a completely new way of just using the models that just wasn't possible like a couple of even weeks ago.

It depends what you're solving for. I think we have a high bar for aesthetics and quality and cinematic outputs and professional outputs. I want to make sure we can keep raising the bar. Part of it is you have to have the right feedback loops for training. And so we have a studio team, a creative team in-house, working right next to a research team. That's how you improve the models on a qualitative kind of aspect. I think if you aim very high for quality, it would unlock a lot of other use cases that are just downstream of that.

I don't think there's one single answer. You have to prioritize for that.

I actually wrote a long post about this very recently, but there are a couple of things. The first one is setting the right expectations for people. I think, as you were saying before, most of people's experience with AI these days, if you look at a macro level, has been with chatbots. And the way you interact with a chatbot is you give the system one prompt, one answer, and you expect one answer back. And it needs to be true, and it needs to be good, and it needs to be like one single thing that I do, right?

I think if you're completely new to creative AI or using AI to make images or videos, you might come with a very similar expectation, being like, I have this incredibly creative thing in my head, which is a complex scene that I only can visualize internally. I'm going to go into this software, I'm going to type the words that I think describe what I'm seeing, and then once it's out, I'm going to be extremely frustrated because it doesn't match what I had in my head.

And the conclusion is, therefore, that it doesn't work. And so for me, it's like you're watching a Christopher Nolan movie and you realize he used a camera for that. You go and buy the same camera, and you press the button to record, and you watch the output, and you're like, those two things don't compare.

Right, exactly. It's not me, it's the camera. The camera doesn't work. I make that kind of comparison because I think ultimately this is a creative medium that requires you to experiment, spend time understanding how it works. If the assumption is you're going to come and make a film by pressing a button on a camera, you're not going to understand how cameras work. If your expectation is to come to any AI creative software and type one word and get exactly what you want, you're not fully understanding the extent of how they work.

And so part of it is just helping people manage their expectations and understand how things work in this new world, that some of the things that you're expecting might not actually happen the way you expect them. It just works differently. And it's just a learning curve. For the first couple of times, it will take you time to adjust until you understand how it works, and you start realizing the potential of it, and you start understanding how you can bring it in.

And I think managing expectations is probably the most important thing for people who are totally new to the field.

So I think there's a lot of things that have worked. But I think this field is still very nascent, and there's so many things you can solve for. I think image generation is not fully solved, but it's made a lot of progress. I think most of the tasks that people thought were going to take many years to solve are instruction-based prompting or instruction-based generations. All of the things that are very specific and detailed, I think models are getting really good at.

On the video side, I would say long consistency of scenes, or being able to cut and have consistent characters within the same generations. Those things are also making a lot of progress. I think real time is getting closer and closer, and I think it will be perhaps the sole focus of many companies over the next couple of months. How do you make sure inference happens on a real-time basis, like you can do with language models these days? You can have a conversation with an assistant. You're going to get to that for video very soon as well.

I think there are many things that haven't yet been solved, and there are things that are getting better. Like consistency overall is getting extremely good. But I think my belief has always been that, as a field, we solve rendering first. We are able to show and create incredibly consistent videos and images. We haven't yet solved control, which is how do you make sure the model is going to create the thing that you want to create. And control has only started to happen over the last couple of years, months even.

So there's a lot more focus on control.

I think it was just time. I think people know if you believe something to be true, it's just a matter of time until it worked. I think for a long time, all of those cultural moments of the six fingers and Will Smith eating spaghetti, I think for me, those are focusing on a specific moment in time and trying to extrapolate from that towards the future, considering that nothing else will change. And I think we had just a different perspective, being like, yeah, those things are imperfect, but you're not extrapolating well based on what happened before that, and before that, and before that.

I don't think it was one particular thing. It was more of, I think many in the field have believed some things are going to be true, some things will scale really well, and building the infrastructure to get there is perhaps the thing that takes you the longest. Like, training a model is not trivial. It takes time. It takes time to do the right captioning, the right annotation, the right infrastructure around it, the right testing. But those things are coming. I mean, they will be solved over time.

There's a lot of things that we learn at every model that you build. You learn something around what works and what doesn't. And I think a model is not just like one single idea. It's a combination of different ideas, from how you caption, how you do training, how you test, how you benchmark, how you do different parts of a model. If you swap specific architectures on the encoder or the decoder, there's many parts of model building that, for me, are more like an art than a science that you're going to learn just by shipping one model, then shipping another model, shipping another model.

Gen-4 has a lot of the things that we've learned over the last three generations of models, and the next generation of models that we're going to be releasing are also going to have a lot of the learnings and the things that worked in Aleph and in Gen-4. And I think that should continue to be the case, where it's less about one single algorithmic innovation that's going to change the entire field and more about how do you make sure that those pieces are set correctly?

Because I think ultimately it's a complex puzzle that has many different parts, and you just need to know which ones are working and which ones are not.

A couple of months.

It depends on the standards that you have and how big the model is, and yeah, there's a lot of—

It takes a couple of months. So the first model that we ever released, I think, was Gen-1 video-wise. It was the first model that was ever out publicly and commercially. It was around 2023. I know that because it was on the front page of The New York Times. Big deal at the time. So 2023 was Gen-1, and now we're in Gen-4, Gen-5 almost. Five years or so? Five years. And so at the beginning it was, like, every 12 months, then I think every eight months.

Then by now, I think things are getting to a point where you can release new pre-trained baseline models every couple of months. I think partially, again, it has to do with the infrastructure, with the amount of work that you—it's hard to see because when people judge a model, they just judge the output. And I think it's a fair assumption that you're judging what you see, but it's very hard to understand what went into the model itself and how good that infrastructure knowledge of the organization can be used to ship another model and another model and another model.

And I think that, for me, is the most valuable part of Runway. It's not a model that we put out, because models will completely change every now and then. It's the organizational knowledge and the infrastructure that it takes to ship a model like that. And if you're good at that, you're going to start shipping them much faster than before, which happens to be the case for us.

We started building pretty much, I would say, almost everything from scratch. And so we've spent a lot of time building really good research workflows and tooling for researchers. And those are the things that are hard to measure and see because there is no immediate output or no immediate value yet if you're building and spending time on those. But I think if you make the right bets on the way you manage your data, the way you manage your cluster, the way you do deployments, the way you do research and training jobs, and all of that infrastructure knowledge we build internally, in some cases, I'm pretty sure if we take some of those internal tools and we make them products, they'll be successful products on their own.

But now, a lot of it has to do with knowing exactly why you're building those kinds of things. And in some cases now we've managed to buy some stuff, like we don't have to build everything from scratch. There's enough knowledge in building those from scratch and knowing why those things work and why others don't work. I think what we can share the most is a lot of what we build is very much custom-built. And I think that's an edge if you can afford to do it.

And I think we've managed to afford to do it because we just started.

Yeah, so we don't disclose what data we use, but we've done some announcements, some partnerships around data. We have one with Lionsgate. We announced another one with Getty Images. And so we have our own internal teams that are collecting data. I think quantity matters a lot, but also quality. So making sure you can curate the right data. Garbage in, garbage out. So if you just put a lot of garbage into the model, you're going to get a lot of bad stuff into it.

But I think quality then goes back to the question you asked me before. It's like, what's good? In art or in video or in filmmaking or in any artistic endeavor, there's no such thing as a right or wrong answer. There isn't, like there is in a chatbot or in a search engine. And so a lot has to do with just training the right eye to select and curate the data itself. And so data for us, more than quantity, is a lot of the quality component to it.

It is. I think it's becoming more of a thing, I would say. It still has its challenges, mostly to generate diversity of data, but definitely something we're exploring.

I do feel that depending on what your goal is, you might just choose whatever is the best outcome for your mission. And I think for us, we know there's a lot to be built, there's a lot of product momentum, that research and product have to work really closely. I don't think open source necessarily gives you that level of control. It has other benefits, it has other things that I think are extremely valuable, and I think there's a lot yet to be built.

And we've open-sourced a lot of stuff before, but I think for where we are right now, we'll probably continue to build models just internally.

Yeah, I think we have it. It's more about deploying it and the unit economics around it. I think you're going to start seeing real-time becoming more of an interesting use case than things that you can see.

Correct. You start with, let's say, a reference image or reference video or a prompt, and you're free to basically navigate this world openly as you wish. And I think what I mean when I say this is a new medium is, the moment you're able to do that—and if you've ever tried it, and very few people in the world have tried it—it doesn't feel like anything you've experienced before. Because if you think about linear media, films and videos and ads, it's the same video everyone watches and it's the same experience.

It's the same sequence of actions happening over and over again. So that's why it's linear media. There's nonlinear media, like games, and in the nonlinear world, you still have instructions being built in. There's worlds and parameters around how things are supposed to happen. And so the rules of the system are already baked in, and you're kind of exploring something that someone already created. Now, this is different because you're starting from something, let's say an image or a video or an initial starting point.

And then if you extrapolate where things are going to go, you might be able to just navigate and move around that world freely, and there's no rules to the world but the rules that you want to have in the world. And that in itself, for me, it's not a film, first of all. It might have some of the qualities of a film, because you might choose, like, the story and you want to follow the story, but it's also not a game entirely, because it doesn't follow the rules and the instructions that we know of games.

So what is it? I don't know. I know it's just different. It just feels and tastes and smells different. And I think real time is probably one of the things that will unlock many other new use cases that just were never imaginable for people before.

No, no, I think a lot of people have come to some sort of the same conclusion of like, "Oh, you can personalize your films or shows." I don't think that's necessarily wrong, although I do think that that's a way of looking at this new medium with the lens of what we know.

So it might not be the case that that's something people want to experience. And that doesn't mean that films are going to go away and you're not going to see shows anymore. It just might happen to be the case that this is a different experience altogether, and there might still be Netflix shows in the ways that we know them, but then there's an experience that you're having on the side that could be inspired loosely by the story that you watch in a linear way. You can have both; it doesn't have to necessarily replace.

And I think when people get angry or perhaps mad, they might be interpreting this new way as replacing the entire old thing. And I'm like, no, it's not gonna replace it. You're still gonna have Guillermo del Toro building the film for you, but you can have another thing on the side. And that other thing might be just a different experience. It might not be for everyone, that's fine. Not everyone in the world plays games, and it's totally fine, but you can still have both.

And you're going to get to a point where you can customize experiences in the way that you're describing. I'm just not sure that they're going to follow the same rules and patterns of films, where you have actors and scenes and sequences. I don't think that's probably going to happen.

So from a business side, I think our goal is to help people tell stories. Ultimately, I think storytelling can take different forms and shapes. Today, it happens to be the case, again, the most obvious one is the people who make storytelling for a living: all the studios, all the agencies, all the media companies, the brands. Most of, if not all of, our adoption is just very organic. We have people coming to the platform. Again, it started with a very small set of subcultures and sub-users in very small niche communities, and then started to grow from there.

The brand is now, I would say, and the product is known by many, mostly because you've seen it somewhere else. Because someone showed it to you, or you saw a video of it, and you experience it, and you share it, you start using it. We have hundreds of people and thousands of people just making videos for their own enjoyment. It's an audience of one. You're just making it for yourself. There's value there. I think there's a lot of value in making experiences that are just for you.

But then you probably work professionally somewhere else, and you bring Runway to your company. I think having your users and your customers be your sales force, in a way, it's great and it's hard, but it's great because it helps you just go into many different places that otherwise would be very hard to get into. That's how we've gotten into all these studios. It's not us trying to deeply sell to them. It's like they reach out being like, hey, we have this use case, or we want to use it because our team was using it.

And we started to see that in other industries, like architects, same thing. There's some overlap between VFX and architecture software. And for some, it became interesting to experiment with Runway, and it became a thing. And then you started kind of growing like that.

Probably 99% inbound.

I think we will. I think we realize the business is in a position where we need to make sure that we can show this to everyone, even if you haven't seen it before. I was traveling, I came through customs a couple days ago, and the officer knew about Runway, and on the visa it said Runway. He's like, oh, Runway. That's great, that's amazing. But we have all those stories where people now have heard about it, and I think we need to make sure that there's many other people in the world—I would say most of the world out there—who just haven't heard about this.

And you're going to get there by just doing everything you need to do.

The best option is just Unlimited. Get Unlimited. Unlimited is a plan that allows you to generate as much as you want. It's like, what, $79? And you can generate everything you want. You can pay for credits if you want faster generations. And for enterprises, it depends. We have people who are using the API to generate thousands of videos, and in that case, we charge you per request. And so it's very flexible, since we do everything. We do model training, deployment, inference, distillation, optimization, product.

Then we can manage to change everything from that stack that they want to change. I think one of the things that I would say in AI these days is tough if you're just on a particular upper layer is the margins, because you don't control the rest. You're just basically giving the value to whoever built the model. I think there are very few companies out there that can do this full-stack approach. You train the models, you deploy them, you do the inference.

And so for us, in some cases, we might charge you differently depending on which part of the stack you want. If you just want the API, we can charge you this. If you want the product, we can charge you this. But I think overall, my general sense is that prices will continue to go down, mostly because compute will continue to go down.

Of course there's cost. But I would say most of the AI labs' gross margins these days for companies to build are between, what, 40% to 70%? I think eventually you get to best-in-SaaS margins, like 80%, 90% over time. And I think most of those companies who are in those ranges are the ones who build the models themselves. If there's a dollar that comes in and you can get that dollar out for the entire stack, then you're benefiting from it.

If you're switching that dollar or giving it to someone else, then in some cases you might not be owners of your destiny in a way. And I think we're pretty much still owners of our destiny.

It's great. Again, when you create a market, an industry, if it's interesting enough, you're going to have the best companies try to follow you. And if you're scared because of that, then I don't think you have the guts to continue leading it. When we started, no one cared. We showed a way, we showed a path. Now others have followed, and I think that's a great validation. It's a great sign. I still believe the companies who are going to win here are the companies who are obsessed around the problem at hand, who are obsessed around not only catching up, but leading the way.

And I think that just requires a completely cultural set of approaches on how you build both product and research. Speed is of the utmost importance these days. It needs to be very fast. You need to learn a lot. And I think for us, it will continue to be the case. We've seen competition come and go for the last six years. Every new year is a new company people ask me about, and it's like, what do you think about this? But, Cristóbal, they have so many PhDs on their team and they're so well-funded. Great.

It's going to move the field forward if they succeed. If not, we'll continue building what we built. And I think that happens to be the case year over year. I'm confident that we'll continue to lead the way.

Be very open-minded. I think if you have a very consistent and particular way of thinking about how the world has worked, I think that's probably not going to adjust well to change. And I think diffusion of technology has changed. Technology has changed the world, of course, all the time. I argue that art is the history of technology, and that will continue to be the case over time. But it used to be the case that we had more time to adjust.

We have sometimes, like, years and decades to adjust. And so the media world had years and decades to adjust to streaming and digital content. I don't think people have now realized that we don't have time to adjust. There are many companies that I've spoken with over the last couple of years that thought that all that what's happening today was kind of supposed to happen in the next 10 years. And if your whole strategy has been waiting and seeing, I think you're not going to make it.

And I think that happens also at a personal level. If you're just waiting and seeing from the sidelines, you're going to miss out. And I think there's nothing preventing you from experimenting, trying new things, even if it's not perfect. I think models have so much value these days in all sorts of domains. So be very open-minded and willing to question the essence of a lot of the things that you know are true, because I think most of them will change.

It's funny, someone asked me that same question a couple of weeks ago. And I think the way I thought about it is, we train models at Runway. So a model is basically an algorithm that learns about data, learns the patterns in the data, and then creates something based on that data. And then if you're good at training models, you can add more data, change the outputs, and keep doing that all the time. I think of Runway as the organization, as a model itself, where there's data in the world, and data might be markets or technology, competition, or talent.

There's data happening. You feed that data to your organization that happens to be the people and the knowledge between the people, and then an output comes out of it. And then what you need to do is take that output, put it back as data that you understand and mix with everything else, and you keep doing that all the time. And I think adjusting the weights of that model is the most important thing AI can do right now, which is: how do you make sure that as new data comes in and outputs keep changing, this model keeps growing and learning and becoming better?

And I think part of it is, sometimes in AI land, you start with random weights. You start with no knowledge of, no understanding of what the model does until you start training it. And then suddenly something comes out and you think, what works? I like to think of an organization that operates in a very similar way. You're constantly learning all the time. You're a system that keeps on learning. The moment you stop learning is the moment you stop pretty much growing.

So perhaps it's a self-recurrent answer, but yeah, I think companies should operate as AI models as well.

Five years? That's like five decades in AI.

We started the company because I was just having too much fun with my co-founders at school.

Really. I'm 100% honest here. We finished school, and we just loved working together. And so we thought that starting a company might be the easiest way to keep having fun and learning about this. And I think I still—what I enjoy the most about Runway is I come to the office now, we're 100 people, and sometimes I sit with the brightest minds in a particular field, and I'm learning a lot. It's just so fun. I think success for me is, like, we'll keep doing that for many more years.

And in the process of doing that, you're going to change the world somehow. You're going to either make the world more creative, you're going to help someone make a story that they couldn't do before. You're going to change how people learn, how people see the world. But most importantly, the organization itself is having fun and enjoying doing it because they care. I think care—if you fundamentally don't care about what you're working on, you're not going to get very far.

I think we care maybe too much. And so in the next five to 10 years, I want to keep on working on stuff that I care deeply about. And then the consequence of that is that you make great stuff and great products. The moment you stop caring is the moment nothing really works.

Thank you.