Runway’s WorldPrompt and the Engineering of Real-Time Worlds

Early days of Runway

0:00Okay, we're here with Anastasis from Runway, with me and Vibu in the studio. Welcome. Good to be here. Congrats on all your success and progress with Runway. You're opening offices all over the world. Did you envision this when you first started out? Not quite. I think even when we started, we had this idea that, you know, it was more a matter of when, not if. We were seeing the early generative models of 2016, 2017, and just extrapolating.

0:30Assuming, you know, resolution quality increases predictably over time, there's going to be a point where most of content will be generated. And that was maybe the initial thesis of Runway was, we will need, as a result of those generative models, rethink how creative tools are made. And as we built out the research behind our generative models, it then became clear that they were useful far beyond that as well. And it is more obvious now with, like, the real world stuff and the world models that we'll talk about later.

Background in art and simulation

1:02I'm just kind of curious how you go from a background in, like, ZocDoc and, you know, computer vision into Runway. Like, take us back to that early conversations with Chris and, you know, whoever else is on your founding team. I was always split into those two worlds. One was the, you know, I had my own art practice. I was making a lot of interactive art, I think, for a long time. And then on the other side, I was working in startups. And I was working as a metal engineer, as a backend engineer at different companies.

1:32I've always been interested in coding and computation and especially interested in simulation and brought it back into my early artwork as well. And at the same time, I was interested in... The personal site has a few, right? Is there one that we should pull up? Just in case there's something that's like... I just like to go down memory lane. Okay, what is this? So this was a project that I made, I think, back in 2015, where I built this software that would give voice instructions to people in a gallery space.

2:02So it would basically coordinate interactions between people. And so it would first give you an identity, like you're an architect, you're 30 years old, and you like sports. And then it would match you with another person and you have this completely generated interaction. Obviously, language models were not quite there at the time. And so it was a mix of some templates and some, like some Markov chain generated text. And it would just completely simulate this small talk conversations between everyone in the gallery space.

2:36So I was always very fascinated on the one end with generative models and like the early machine learning work that was being at the time. But at the same time, there was a separate thread of simulation and what it means. Like, what can we learn about humans by creating those very simple models of their interactions and their behavior? Did you generate the prompts or, you know, the 30-year-old, whatever? Was it you generating them? How'd you? Exactly. So the program would just generate those from, a lot of it would be kind of Mad Lib style of just, you know, you have lists of different professions, lists of different personality types, lists of different ages, things like that.

3:18And then it would just combine those things together. And then maybe the next project we go is Uncanny Valley, Uncanny Road, which was one of the first projects that we built with one of my two co-founders, Chris. This was taking Pix2Pix HD, which was one of the early image-to-image models that NVIDIA released back in 2016 or 2017. And it was a model that would take a semantic map of a scene and then generate a photorealistic, let's call it, output.

3:54Obviously, very, very early days. So it was not very high fidelity outputs, but I think it was the first image-generation model that we generate at 1K resolution. And it was all trained on self-driving data sets. So the semantic categories it would support were only, you know, things you would encounter on the road. So it would be pedestrians, traffic signs, bikes, stoplights. And so that was actually one of our first indications that we built this and people were making all this like very surreal imagery of a million pedestrians or a million traffic signs or like gigantic humans.

4:32And it was an indication that you could take a model that was trained on this very boring data set, essentially, of like not that many interesting things happen when you're on the road. And then you can repurpose it and go very out of distribution and make something that was artistically compelling. And that was, it's a summary of the thesis of Runway in some ways. Like you can take the safe generative models and if you look at them from another direction, if you build interesting tools around them and you give them to artists, they're going to do things that you don't expect.

5:02Very cool. I like the, you know, UX of it, basically. You're just given an empty canvas, drag whatever, do whatever. And then the other one, like you see everyone with wired headphones, you know, like that's a sign that it's very easy. The Apple original ads.

Scaling up generative media

5:18Yeah, yeah, yeah. Take us to today. You've been doing this for seven years at Runway. How have we got to this, you know, like how did we go from driving simulator data to all this? And you kind of cover the whole stack of generative media, you know? I mean, interestingly, we're almost back in, you know, we're full circle where we're now applying our models and kind of beyond creative tools into real world scenarios. But it was a, it was a long journey. It was very early on. We realized the, you know, the first version of Runway was a way to easily use the, all the open source models of the day.

5:55Things like Pix2Pix to, and give them to artists. That was, that was the initial idea. Is those models are too difficult to use if you're not a machine learning engineer. Like what happens when you give them to artists. Very quickly, we realized we needed to build a research org inside of Runway. And that happened maybe on year one. And a lot of the mandate there was the image generation most of the time, the video generation most of the time, or there were barely any video generation all the time. But they were not quite there where they could be productionized and brought into tools that would be part of creative workflows.

6:27So we need to push the frontier of the research. And so maybe the, the first four years of Runway research was almost happening on the background until there was a moment in 2022 with latent diffusion, with DALI2, where, you know, there was that step function change. And you guys maybe remember. I started LainSpace because of basically latent diffusion and stable diffusion. Because I was like, wow, this is not only like feasible, it is actually doable on consumer hardware.

7:01Exactly. I think the Delta is also huge. Like I learned pics to pics. Like this was intro to ML, the TensorFlow, like Jupyter, Google collab notebooks are like this. And then you have a sudden step function change, you know, with diffusion and whatnot. Any other ones since that, like there were clear examples of what early diffusion were to get to here? Any other changes in key technology research? Between 2018 and when we started in 2022. So one of the early work that we did in Runway was solving segmentation, image and video segmentation.

7:36It was a very important problem because most VFX involves essentially separating subjects. Yeah, rotoscoping. Extremely manual process. Nobody enjoys doing that. And so a lot of the early days of Runway was building this tool that was called Greenscreen. And it was for a long time the main thing that people were using Runway for. It ended up being used in everything, everywhere, all at once and a bunch of other kind of high visibility films and series. But that was essentially Runway for a long time, was a post-production tool until Light and Diffusion, Gen 1, Gen 2 happened.

8:11Cool. I mean, let's go past that moment. You've come a long way. Then you started releasing your own models. Maybe describe that journey as well. Yeah. So we go to the other point, yeah, in kind of mid-2022 when it became clear that we're doing research at a fairly small scale of compute. And it became clear that, like, scaling laws would apply to image and video gen in the same way that we're applying to language generation. So we made a big bet. And I think, so at the time, we signed this deal to build a cluster of 1,000 A100s, which at the time we were a Series B startup.

8:49That was almost, you know, slightly irrational decision maybe, but we really believed that if we trained a video model at the large scale, we would get, like, a great model at the end. And at the time, the goal, you know, the goal, or we set the goal around fall of 2022 of what is, what does the latent effusion, stable effusion moment look like for video? You know, and at the time, the best model of the time was called cog video. It was one of the early kind of video models.

9:20It was very 256 by 256 resolution, very, not very high quality. And so we decided we're going to build out this cluster and we're going to just invest in, like, in building out our own video model. It became clear as we're training Gen 1 that it was difficult to get to fully. We wanted to build text to video, but it became clear to us that an easier starting point would be to start from video to video. Because when you have a stronger conditioning, it's basically an easier problem to restyle as an existing video versus generate a video from scratch.

9:57And so we released Gen 1 first back in, you know, it was January of 2023. It's just a fun visual podcast, honestly. Like, we can see February 2023, what was the state of stuff? I mean, it's so interesting because at the time when you see those results, you think this is, this is so incredible. And this is, like, it's almost like image generation or video generation is solved. And then you look back a few years after and it's like, it's obviously, it's just like you get used to results very quickly with those models.

10:30But at the time when we started seeing those results, it was, it was, you know, it felt quite, quite incredible. And the level of, like, quality you could get. And so the Gen 1 was a depth conditioned video model. So it would turn, it would, it would take an input video, it would predict, it would, you would first convert it into the depth map. And then we would generate pixels with a, with a light and diffusion model.

11:00Yeah, very effective. I didn't realize how distracting the blog posts would be, sorry. Yeah, but one of my favorite examples actually of the, on those, on Gen 1 was both, if you, if you go up to mode three or mode two, there was this storyboard use case where people would make, would basically, um. You could mess around with the room. Make a city out of books or out of boxes. And then they would, they would kind of shoot a video with their phone and then translate it into a, for the realistic output. There was all these ways in which those models were starting to be used for storyboarding and also for really, uh, and then if you, if you go to mode four, like of taking kind of untextured 3D scenes and then turning them into photorealistic output.

11:42So we saw a lot of use cases early on where people that were familiar, you know, where power VFX editors would just take a, a blender, a render, and then they would get translated in with Gen 1 or create a scene in Unity. And then take a, capture a video of it and then, and then translate into, you know, restylize it. So I still, I still think video to video is powerful. I think we, we had a recent video to video model as well. And it's one of my favorite ways of using, using those models is essentially using them to, to use ground truth video as like the initial inspiration and then translate into, into different styles or different outputs.

The stable diffusion story

12:23I think we're going to go into like the rest of runway and catch people up to speed today. I did want to cover the, let's call it the, the stable diffusion controversy or, you know, what, what happened with stability AI, whatever, you know, I think there was a sort of two sides of the story. I think there's part of that is a normal thing of like people, you know, join and leave companies, but what is the, you know, retrospective now that, you know, there's been some years behind it. Yeah, it's a very, it's a very long story to go, to go into, I think.

12:53Which I remember you actually wrote a really long post about. We would probably cover the whole hour to, to, to, to, to go into it in a, in more detail. But essentially, you know, there was the latent diffusion paper that came in, I think that was at the, at the end of 2021. And then Patrick Esser, who was one, one of the researchers behind latent diffusion, and he worked at Runway at the time, he built latent diffusion collaboration with Robin Rombach and a few other folks back in the, in a convys, which was a, a lab.

13:26Like a research group. Yeah. Yeah. And, uh, after releasing the early latent diffusion model, they, um, essentially there were, you know, the goal was to keep working on versions of the model, kind of scale it up, uh, incorporate new data, incorporate new tasks. And stable diffusion was basically the same model, but trained on more compute. And then with a few more tricks, like a classifier free guidance paper came at some point, I think at the early 2022. Which like was a big prompting improvement. Yeah. You know, that improved results, I was trained on better data.

13:59So like a, the aesthetic subset of, uh, of Lion, but it was effectively, you know, the same, uh, the same underlying architecture. And there was that big training run, uh, that, uh, happened on Stability's cluster, uh, stability kind of financed, uh, financed that run. And looking back at that story, I think it was the work to build and train that model was, was done. It was a, it was a research project. It was done as part of the continuation of the latent diffusion work. It's then, I think the model became very successful and it, um, I think there were the, and, and I think as, as a result of it's, uh, a success, uh, their companies try to, uh, figure out the commercialization path for it.

14:43But for us, it, it, it was very important that we try to, you know, we, we, we make sure that we, it was meant to be an open source research project. And so the, the, we decided that we should continue releasing versions of it, uh, since that was kind of the original, the original goal of Stable Diffusion and that led to releasing Stable Diffusion 1.5. There was maybe a day of, uh, a bit of, uh, miscommunication there, but ultimately that was resolved very quickly within, within hours. Um, so yeah, there was a nice, I just wanted to, you know, you have to, you're actually one of the main players in, in that sort of journey.

15:18And so it's nice to hear from the source of like what happened. Yeah. Yeah. I mean, uh, I, I think it's all, it's all in the past now, I would say. Um, and like both companies, you know, Stability took its own path, Runway took its own path. There's still, I mean, James Cameron is backing the new Stability, whatever they're doing with the Hollywood studios. Right. I don't know what they are doing. I think one thing that impresses me and I'm happy to move on is that back in the, that time, let's say like 2021, 2022, there was this community of people that you were involved in that was researching all this stuff.

15:52Right. And like from everyone I talked to who was active then, it seemed like it was fairly obvious that somebody would do the hero training run that would produce Stable Diffusion. And so I guess the question is, you know, like you had the, you had, you were, you had made investments, you, you, you, you had the foresight. Is it accurate to say like, that is reflective of like what people were thinking at the time? Or was it still very much like, well, we'll use it as like a post-production tool or something. I don't know. You know, like, like where, where, where in the sentiments were we that maybe you can sort of think back to like what, what the community was like back then?

16:27I reminisce and I, um, I, I think of very fondly those early years from like 2018 to 2022, because it was a very small community that, as you said, were very convinced that this was going to be a big thing. And at the time, you know, anyone who, because it was such a small circle and everyone who would like be part of that circle and like make projects with it would, you know, immediately kind of get, you know, go viral. Uh, so like, you don't even know who they are, right? They're just, there's some name on the, you know, a GitHub dog hugging face somewhere.

16:59Exactly. Yeah. So, uh, so I remember one of the first big viral moments of creative AI was, uh, there was the neural style transfer paper. Uh, uh, something dreaming, uh, I think it was called neural style transfer. There was also deep dream, the puppy slide, which was, uh, also, also really cool. Uh, but, uh, there was this project that, uh, Gene Cogan, who was an early advisor of runway and one of those, uh, Marketing guys, big, uh, creative, uh, creative AI, uh, folks, he literally just like shot a video of himself taking the New York subway kind of and going over at the Williamsburg Ridge and then stylized it with, uh, I think in the, in the style of Van Gogh or like one, one, a painter.

17:44And, and that was like, at the, at the time that was like so, so cool and it went viral and it was completely revelation to people that you could do this with generative models. And that was only, you know, it was less than, it was maybe 10 years ago. So just like as an indication of like how, how quick like things have gone. It's pretty crazy. Like even since then you've kind of got people at every level of the stack, you've got devs, creatives, artists, hobbyists, you've got everyone using it. And for people that tried stuff early, they'll remember how hard it was to use regular diffusion, right?

18:16Like nowadays you can use your favorite chat GPT image gen or whatever, give a sentence, get a beautiful output. But diffusion was like, you know, the whole ultra HD, 4k high resolution, like prompting these things was very different. Uh, anything you learned on the tooling side, like from the offerings you guys have now. So like creatives, devs, uh, you really took the research and brought it to everyone to use. Um, anything interesting there to share?

Building video model control

18:44We had to build the entire model serving infrastructure for video diffusion models. There was no, nothing else, uh, already. Like, cause we had gen two was the first text to video model, I think out in the, in the market. So many things that we learned over time. Um, I think the, I think the biggest one was like, we, you know, it was very clear early on that text to video was not going to be the, the answer. Like you, like people wanted a lot more control than that. And so we invested in like control building on top of those models very quickly.

19:14You know, how, how do you use the camera trajectory as control? How do you use an initial input frame as control? So, so that was a very early learning for us. It would text to video was, um, like gen two was an amazing, you know, step function improvement in the quality of video models. But it was used much more in an exploratory way because there was nothing to ground it to. There was no reference that you could bring into it. There was no, you couldn't really control the camera motion. You couldn't control the object motion.

19:46And so the first year in 2023 was really all about what are all the interesting ways in which we can condition those models. And it was a lot of just post-training grants on top of the base model to figure out like what, how do people actually want to control them? And so there was like this quick succession of the, we, uh, it was called motion brush. We suppose you could like, you could basically draw arrows and dictate where things should move in the scene. There was camera control that was, you could just describe like how you want the camera. to move in the scene. And because we work with filmmakers from kind of the, most of the history of runway, we immediately got this feedback and got this, you know, decided that this was worth investing in.

20:24And so controllability became a big theme. I think very early on as we're building, as we're building those models, something fun that I haven't really talked about too much was just how gen two came to be out of gen one. Um, so, so it was a bit strange cause we announced gen two, two months after gen one and, um, it was before gen one was even generally available, but gen, gen one was a depth to video model. So it would take a depth map and it would convert it into RGB. Um, and we couldn't get, you know, text or image to video to work directly.

21:00And that's why we started from depth to video. Um, but, uh, you know, and we, we had discussions of like, okay, we need to spend the next six months actually investing in text to video, maybe increasing the compute scale or the model scale, like train a larger model. And I, I had this weekend project idea, which was, what if I take a model that, uh, starts from text input and converts to depth maps and then, and then use gen one to convert the depth maps into RGB. Uh, and so, and so gen two was basically that.

21:32Oh, but it looks good. And it worked pretty well. I mean, there, there were, you know, if you, with the knowledge that it has this like two stage pipeline. You can tell in some cases that the structure of the video looks a bit off because you had to generate the depth first before you go into, into the output video, but it worked and it allowed us to bring, bring this to, to our, to users very quickly. But it's actually now it's interesting because like people are coming back to this almost two stage approach.

22:07Like if you look at the, the Reeve text to image model that came a few months ago, it had this, um, this planner model that would generate bounding boxes before it fed that into a diffusion transformer. Yeah. Ideogram also the same day. I remember that was very strange that both of them came out the same day with the same exact innovation. It's a small community. I think. I'm like, this is like, this is completely coincidental, right? People talk.

22:34So, so yeah, there's, there's definitely something into this approach and obviously now like every single video, like video generation model in production uses a complex prompt completion pipeline under the hood. I think that's no secret that there is. Humans are terrible at prompting. Um, I think across the board, but yeah, I think like the original Sora one blog post even told you that what happens after your input is re rewriting your prompt. It's much more descriptive about what you would want. Exactly. Yeah. Um, and there was the DALI three paper beforehand that, uh, was the kind of the, the first public, uh, description of the fact that synthetic captions and really detailed captions work really well.

23:16And then Sora kind of built, uh, built on that. And yeah, so it was 2023, we were kind of releasing all these updates to gen two, like the camera control motion brush. And there was actually something very interesting about camera control because it, it was the first time that he felt that instead of like you were creating video, you were creating a short video, you were actually navigating inside the world. And I think camera control was maybe the seed of some of the ideas that we had around the world models and really opening up that research direction.

23:46And we realized, you know, it was this era and this series of, you know, gen one and gen two models really proved to ourselves. Yeah. This is, this is the, uh, so this is not the original camera control. This was the update to camera control on top of gen three, but yeah, I think it made those models usable to, to filmmakers. Uh, I would say the, so camera, camera control was very, uh, was, was very popular. And so we realized, you know, there is one way of seeing those models, which is, you know, you're just as a, as content creation machines.

24:20And that is the other way, which is you're, as you're predicting video in order to predict video, well, you need to simulate the world in an increasing and increasing capacity. And if scaling laws apply on video, just like they apply on language models, then as we scale the computer, we put into those models, then they're going to be able to simulate physics. They're going to be able to simulate human actions and dynamics increasingly well and predictably well. So that was the thesis about around our efforts on the world models.

World models and physics simulation

24:51And we spin up this research group to just focus on, on the world models and how, how do we turn the video generation models that we're building into something broader and something that would be useful beyond, uh, also content creation as well. And that was roughly when? Yeah. So that was in, in late 2023. Interesting. You know, like, I think, uh, a lot of people have been saying a lot of video gen model companies have all pivoted to world models these days, but like, you know, 2023, you're posting it, uh, one.

25:21It's debatable whether it's a pivot, like arguably that's what you always had to do anyway, right? In a way, an expansion of the applications of the models as the, as they become more capable. The early signs, it seems like the original models you guys had, guys had, people would say it's very not bitter lesson pilled, right? You're adding, rewriting prompts, you're having all these one-off things, but that's just the state of the tech as it was versus the future of, as you said, you can scale it up as, you know, we can scale up to world models. Yeah. So it just became, and, and, and if you looked at the outputs of Gen 2, it was not, I think it was not obvious to people that this would scale to become a general simulator of the world.

26:01Like you had very limited movement and you had, you know, very low fidelity or resolution, like all these mistakes in human anatomy, like all kinds of limitations, but it was just, you know, uh, the idea was that's just GPT-2 and GPT-2, you know, you can barely generate like coherent sentences. Similar Gen 2 can barely create coherent video, but if you scale it up, you're gonna, there is no reason why it shouldn't work in a way. It's, uh, and I think that was, that's, that's always the mindset of kind of, of runways, like this extrapolation of like, if, you know, like even when we started in 2018 and you looked at the results of the day, you need to look more at the trend of like where we were in 2018 versus when we were at the, you know, when the first gun came out in 2020, uh, 2014 or 2015.

26:52And, uh, you started from like 32 by 32 images of faces. And then by the time in 2018, you could generate, you know, uh, street images at the 1k resolution. And it was kind of the same with world models, very early science of something much bigger. Yeah. I was almost thinking like it's kind of diffusing into focus. Like if you look at our visible output from year to year to year, it looks like a diffusion process itself. Yeah. Especially watching the early, like all the old blog posts, you can really see the choppiness, the details.

27:24Like human civilization starting from random noise and then difty noising into. Yeah. I mean, just run it a hundred years. Yeah. That's how you know you're on track, you know, you're still noising, right? Yeah. I like the way that you guys phrased it when you announced it in June, which is, you had a sort of video essay. The human mind is no longer the centre of AI, our world is, right? Which is, uh, you know, let's, let's call it the past five years of, of LLM based AI is very much like trying to emulate human preferences and human speech. But now that's like mostly solved.

27:55I think that's like some of the context of your essay, which you also wrote around the time. And now it's like the focus of us on modelling the world accurately. Exactly. Yeah. So the way we see it is, uh, there is that, um, that, uh, initial mission statement of deep mind, which is, uh, uh, solve intelligence and then use it to solve everything else. But I, I think it's starting from everything else, uh, could be valuable of like starting from, you know, there is just so much complexity, uh, and detail in the world that in order to, that it's, it's hard to learn directly from just human descriptions of the world.

28:32Like we're assuming that, you know, like language models learn from everything that humans have written about the world, uh, like our own understanding as of, you know, the 2020s. And there is just so much that we don't know. And so much that's not captured by existing texts, uh, about both the, you know, the low level dynamics of the world. Like we're not describing in detail. You know, if, if I tell you to describe like how do you tie your shoes, that's a very difficult thing to, to describe in words. But it's very obvious thing to demonstrate.

29:05And so I think there's been, and there's, you know, more of X paradox, like we're constantly underestimating all the complexity that goes into very like things that we do subconsciously as humans. And we don't even necessarily always have the words to describe them. Um, and so in my mind, the simulating the world and simulating, um, physics, simulating the dynamics of the world has always been kind of underestimated, uh, uh, compared to, uh, we place too much emphasis on the things that are easy to talk about.

29:39Um, but there's just all this complexity and kind of richness of the world that if we just try and train directly on that observational data, instead of training on how people describe the world, we would learn something new that we wouldn't otherwise know. You know, you think that the present architectural paradigm is fine. You don't need like another layer, like a JEPA, you know, like another famous, uh, New York AI leader would say. You know, we're a very pragmatic research lab. If, uh, we have evidence that an approach works better than the approach that we're taking, then we have no qualms to taking it.

30:14We just have seen no indication that video prediction itself doesn't scale. And even if you look now, you know, not just our work, but the work of others, you're seeing in robotics, some of the most promising work, um, starts from video prediction models, and then you adapt them to also be action models, for example. Um, so there is very little evidence that you need something else and that your time is better spent on a novel architectural change compared to improving data and improving the, um, and scaling the current, the current approach.

30:47And so we don't have any indication that, you know, there, there, there, there is that counter argument that, uh, I think there was, uh, a tweet by Jan LeCun a few days ago that, you know, uh, understanding the dynamics of the world is very different than, uh, generating, uh, cute videos. And your answer is no, they're the same thing. My cat videos are the same as understanding physics. Right. Cause if you want to generate, you know, obviously video models can cheat and like they could, you could give like successive, uh, shots of the scene in a way that doesn't require you to actually simulate difficult physics.

31:23There is always like all these different ways in which you can hide the deficiencies of the model. And it's important not to be kind of too tricked by the performance of the current video models. It's easy to, you know, cherry pick examples and, and think that video models are further advanced than they actually are. So there is a lot, a lot more work that we need to do to improve those models. But in my mind, very similar to language and like we've, you know, you go from barely coherent sentences to something that, you know, could hold a conversation with a human to something that could, can operate autonomously for, for a day and like create an entire code basis.

32:02And the main difference, there is obviously some architecture improvements along the way, but the main thing is scale. And so it's the same bet for video. And we have no indications that this is our trading. Like we have benchmarks that we use for measuring the physics of those models. And we see those predictably improve as we scale those models. So there is, if you want to pull up a physics IQ, is one of those benchmarks that measures how well does the model performance, all mechanics or fluid dynamics or optics. I'm curious if you've seen any emergence, any scaling law around this.

32:37Yeah, he's saying there is a scaling law, right? Exactly. So, so, so, so, so the way those, those, those, those models, those benchmarks work is you, you know, the researchers have gone and like captured a few videos that are representative of different physical phenomena, and then you can take the first frame and then pass it through an image to video model and then generate, generate kind of a rollout that shows what should happen next. So you have a ball hanging from the ceiling, and then you use that as input, and then you, you, the model predicts how the, the, the ball should fall on the ground.

33:10And this measures, you know, we have an intuitive understanding of physics. I know, you know, you can imagine what will happen next if I drop this, this bottle. So it's measuring that same intuitive physics understanding of those models. And we've measured that at different model scales and we see, and compute scales, and we see that the score and physics IQ predictably. So it improves, there's other, you know, tricks and techniques that you can make to improve the score even further, but even scale alone helps in, in the, in the model learning better physics. My main sympathy with Jan LeCun is the, uh, Plato's cave allegory, right?

33:46Like you're like learning on the output of a thing, not the internal process of a thing. And it's very, very noisy. And, you know, if only you could observe the internals of a thing, it's hard to observe the internals of a human mind, but you can very much observe, or at least we have a whole branch of science and physics that we're ignoring on how to model physics and, and movement and, uh, you know, gravity and, you know, other interactions. Uh, and we're just like throwing away all of that and just saying, just, just scale data, which is very much the lesson of unsupervised learning, but it feels wrong.

34:19I think that's the main idea. I think the history of machine learning is, uh, at large, it feels wrong. It's a bit of a lesson, right? No, it's, it's, it's the simple answer to that. I guess, how much can you scale? So like, even on, let's say the video generation side, like there's one side of video understanding video generation, are we still going to have tools where it's like, I want to generate two hours, 20 hours. Um, there's a infra way to do it in batches and stitch it together. But like, do we just keep scaling? Do we just continue long generation consistency, all that would scale and like tying it into where we're at now from, we looked at runway two to 4.5.

34:58Like technically what, what advancements have we made to today? And then where do you see things still going? So part of the answer is definitely scale.

Gen 3 and training infrastructure

35:06Um, and that was, we, we learned that lesson in, in, uh, in a big way for, with Gen3. So Gen3 was the model we released the year after, um, like in 2024. That was a few months after Sora was released. Um, so yeah, there's an interesting story of that, that, that, that, that came to be as well. So Gen3 for us was, you know, the, the first time that we need really needed to build, basically, we had to learn all the lessons that the language model world learned in, in, in, in three years, in the span of a few months.

35:36Uh, one of the biggest changes of Sora was using diffusion transformers instead of covenants. So a lot of the early, you know, latent diffusion models were all, um, uh, covenants for the diffusion model part. And the diffusion transformer paper came at some point in 2023, and it basically showed scaling laws for image, uh, diffusion transformers. And we, we realized at that point that we needed to invest in infrastructure for model parallelism, for really scaling, scaling training to larger than, you know, a few billion parameter models.

36:13And we spent maybe the, you know, most of the fall of 2023, building out our infrastructure for distributed training. And we had a lot of false starts and a lot of failure in trying to scale, um, image and video diffusion transformers. And at that point, uh, you know, February, 2024, Sora comes out and the results are very much superior to what Gen2 could produce. There were a lot of, you know, a lot of chatter on Twitter about Runway, Runway's done.

36:45Uh, like there is, uh, there is no way Runway will catch up. And if you remember also OpenAI in the early 2024, it felt very, uh, like it's a, it's a formidable opponent now, but at that point, it, you know, there were on the top of their game, you know, nobody could even get close to them. There was maybe Gemini was just the first version of Gemini had just released. So when OpenAI came with Sora and it was such a big jump of like quality, it gave me, there was like an existential crisis for, for a few hours.

37:17But that, I think the, the amazing thing about Runway and like, I think the, you know, we've been around eight years now, which is almost weird dinosaur in AI. And we had to like, there was a lot of those moments where to learn, adapt very quickly and build out skillset in the team that we didn't have. And so, you know, if you ask anyone, what is their favorite time at Runway that was there during that time, it was that, that push in like three months to get to a model better than, than Sora. Uh, and it, you know, we scale 10 X, the model scale, the, the, the, the, the model size and the, you know, compute that we were training on, uh, we've figured out model parallelism.

37:54We had zero expertise in that. And then we came out with Gen3 during, during that summer. So, uh, so that was a big turning point, I think for the company where the, the research or grew very quickly. And we really started pursuing this vision of the general world model, uh, in, in earnest, I think after, after Gen3 was out. Yeah. I mean, that's the, the amazing thing about building when you're building, there's no stack to, you have to invent everything yourself. You have to be completely full stack, you know, now, now I, now I think like there are inference specialists like Fal or whatever that can help with like, uh, model serving.

38:27And I, I think you guys work with them as well. Uh, but yeah, like it's, you know, at the time it was just, uh, it's very interesting to think. Think about what you do when, when Sora comes out and people are questioning whether your company should still exist. Yeah. And yeah, there was no, you know, there was no VLM of diffusion models that we had to build the whole model serving infrastructure and make things efficient. And a few months after we released Gen3, we released the turbo version, which I think was the first step distilled model in production. There was a whole trend that we covered as well.

38:58Yeah. So that allowed us, uh, actually to serve, uh, to serve the, those models at the, at the larger scale. Cause I think the first version of Gen3 was, was quite, uh, expansive to serve. You know, I think that the whole like trend in like consistency models, lightning and, uh, turbo and all these things somehow didn't really stick around. I don't know if you have any reflections on this because at the time I was like, well, obviously everything should start with a distilled model first. And then you can upscale, right. Then it basically your, your bigger models just turn into fancy upscalers, but like you should always draft with a, a smaller model and faster model.

39:35Right. Because you can get it so quickly, like near real time. Yeah. I would not be so sure to, to say that didn't stick around. I think that's, uh, it's, it's likely to, uh, I mean, that there is a lot of step distilled models that are actively using production. Uh, there is still obviously a gap in quality compared to the, you know, the, uh, non-distilled model. Uh, but in my mind, we're still, you know, there is a two to three year offset from language models. So the things that, um, so it, it just a matter of time before there is better distillation techniques.

40:11Um, you know, we, we use right now, we have a real time model core characters that I think is the largest deployment of real time video models. Um, that's a, a step distilled model and it's actively being used. It's a very specific use case compared to a general video model. It's very cool by the way. This is an avatar. Consistency character. Yeah. So this is a, a talking avatar, uh, model. Uh, we were able to, you know, we, we, we optimized the hell out of it. Uh, and, and it, uh, it generates at 24 FPS and it's a, it's a step distilled autoregressive video model.

40:45So if we look at our world model direction, uh, a big component of it is starting from the bi-directional diffusion that basically generates entire video at once and making autoregressive shows. So you generate one frame or a few frames at a time. Um, so there's a lot that goes into that pipeline of getting to a real time model. Um, it's first, you need to make it into a causal autoregressive model, and then you need to turn it into, you need to do some additional step distillation to get it to actually be real time. Um, and I think that part is actually just, just starting.

41:18Um, I would be very surprised if we're, you know, two years from now, we don't primarily use real time models to me, real time video generation. It's just inevitable that, you know, it has much better user experience. It's much cheaper to serve. And, you know, the quality gap between the base model and the real time model is only going to close as we, uh, as we figure out better, uh, distillation techniques. And we made a lot of progress there internally on maintaining the quality of the base model when we, when we distill them.

41:49How much of this is transferable? So is it the same base model? Like if you're doing diffusion across the whole sequence and you're converting it to step autoregressive distillation, is this like distillation where you still need to train both? You can use the same base and converter. What's that process like to go from regular model to something that's real time on a technical level? So the nice thing about diffusion models is you have, uh, two axes of distillation. So there is the, you can distill to a smaller model, which resembles what you do in LLMs, or you can distill in, in terms of taking less steps, uh, less diffusion steps.

42:25So you could take a model that generates in 50 steps and, and, and generate in four steps and get to, um, you have some performance, uh, degradation, but very often you get comparable, comparable outputs. So you can even take the, the, uh, you know, the large frontier model and distill it with step distillation and get to a real time performance. And that's what we've seen. So, uh, depending on the use case, in some cases, we might also serve with a smaller model, but in a lot of use cases, we actually just use the frontier model and we're able to make it work in real time.

42:59Um, I think this might be a good time to cut over to his laptop to show off some of the real time stuff that you're doing.

Interface world models for software

43:06This is one of the research updates that we, uh, we did recently. Um, so we've been working in, uh, in getting our general world models to, uh, different applications. Uh, one of them that we think is very, uh, is very compelling is using general world models as essentially, uh, an interface, a universal interface to software. This is, uh, uh, a version of our, uh, world model that's called the interface world model. Uh, and the idea is that it, uh, it essentially, uh, replaces, uh, uh, the front end of a software application.

43:43It renders the pixels directly of an interface and it's trained to predict what happens next as a result of, uh, a click or another interaction you have with the interface. So this is all pixels. Uh, it's, there is no HTML, CSS, React that's powering this interface. This is directly at the output of our real time, uh, video generation model. And it takes clicks directly as input.

44:09And drags, click and drag. Right. So it supports, uh, yeah, clicks, it supports drags, uh, it's also for scrolling. Um, and the, the, the amazing thing about this is that you can effectively describe in the prompt, how you want different elements. Like what do you want the behavior of different elements to be? So it's almost your, you can turn, uh, an interface from, uh, you know, markup language description of like an HTML interface. And instead you can just describe the interface.

44:39You know, if I press this button, I expect this to happen. And if I press this button, this should happen. And it's useful. So we believe both for prototyping for like, just testing, like what different interactions would feel like, um, you can also add audio to it. So it's a, uh, uh, video audio generation model. So you get, you essentially can describe both what the visual outcome should be of your click and also what the, if, if there's a sound effect that comes out of it. So we, we believe that's going to be a much more flexible way of building software, just render it.

45:11Just, you know, why, why generate the code that generates the pixels, just generate the pixels directly. Um, uh, uh, it's the end to end philosophy applying, apply to, uh, to front ends. So we think there is a few interesting use case, so you can build creative tools on top of it. Um, we think that, you know, for any kind of use case that involves a lot of exploration or like, uh, educational use case where you want to learn about a new concept and you want some kind of visualization and, uh, and kind of open ed exploration.

45:47We think those, you know, this is a very powerful, uh, approach, you know, you can imagine new, new forms of, uh, design in kind of industrial design software that could emerge as a result of, uh, of, of those models. And this is all, uh, generated in, in kind of in, in, in real time as well. So, um, you, you can, you know, you can build a lot of interesting kind of camera transitions and kind of forms of interaction that are very difficult to, to build.

46:17Otherwise, and one way in which we evaluate this is what if you try to generate the same interface with, uh, with cloud by just, you know, prompting cloud, here's an image reference of my interface that I made in Figma or that I created some, uh, somewhere else, uh, create this particular interaction, which in this case, it's, you know, drag that object, uh, upwards. Um, and beyond it being slower, it's also very difficult to capture some interactions by just fully, uh, with, with just LLMs.

46:51So we think that this, this is likely to be the way that a lot of the future, like software in the future will be created. Uh, and one of the additional benefits is personalization might be a lot easier done with, with those models. Like you can essentially try out different prompts based on who is, who's visiting the interface. You can, uh, more easily, you know, prompt engineer the, the interface to have larger size, um, text for more accessibility reasons, or you can make this, or, or like, if you have a particular study preferences.

47:24So, so we're very excited about this approach is obviously early days, and I think we'll need to, uh, make it more cost-effective as well to serve those models goes, you know, running a real-time video model versus just purely rendering HTML. There's obviously the computational needs are much higher, uh, but we do see a lot of potential in this approach to building kind of front-end interfaces. So we covered this similar thing with Flipbook before with our, uh, Ethan Hurd episode with Grok, uh, video, and I think it's very engaging visually.

47:58I think it's maybe very good for education, but it's, it does sound expensive. I think there's an upper bound to how expensive it will be though, right? Like, uh, you know, the inference costs will go down over time. You'll figure out ways to optimize it. Effectively, when it pauses, you don't, you're not receiving human input. You don't have to generate anything, right? So, yeah, I mean, you could also like, in this case, you have ambient motion, so there is parts of the, uh, of the screen that might, you know, if you, if you're, let's say you want to, um, visit Paris and then you, you get this interface that you allow you to explore, you have people walking or like things happening, but, um, it's, it's an, yeah, it's obviously makes it more expensive because you need to run the model all the time.

48:37Maybe you have some looping mechanisms, so you don't need to do that, but all those things, I think is stuff we need to figure out. Yeah. I think our first consideration is let's make this clearly find some use cases where it's clearly a much more compelling interaction compared to traditional interfaces. And then it's a matter of time before it becomes more cost-effective to serve. Yeah. When it comes to the people walking, you know, I think the approach that makes the most sense to me is basically Moon Lake. I'm like, oh, I keep messing up their name with Chris Manning and Fanyun Sun.

49:08I don't know if you've come across them where they basically map to some kind of game engine. I think it's Unity or something or Godot, Godot. And you can obviously, obviously script some NPC behavior behind that and train on that. Whereas here you can really imagine whatever you want. Like that is a UI, right? Like, and, and it feels like more tractable, I guess, to, uh, create a role model of software that is interactable because we have many of examples of that. And you can, you know, do your fancy RL environment stuff on that. Then it is scaling up to embodied and real world physical use cases.

49:41But this is a nice first step. Or, you know, there's the opposite of you have like 1B models, 350 million parameter language models. It just gets so small that they're just predicting like, you know, fishes moving. I mean, small models are now 120B.

49:56Ultra many on device. But, uh, no, I think it like, it puts it into perspective, at least the car one for me, like the applications, right? The amount of work to do that. Sure, you only make one model year car per year, but applying this, it's also a cost saving to have to manually make all this, right? So it opens up a lot of possibilities too. I'm curious if you extend this out two, three years. So where do you see things going even further, you know? So effectively the, the, the end game of something like interface world models is you have a fully neural operating system.

50:33So I think, uh, uh, Andrej Karpathic has written about that quite, quite a, quite a while back, but it's, uh, you know, I think to me it's, it's a bit, um, it's a bit odd that, you know, we have, for example, with an interaction with, with an LLM of today, LLM of today, you have this LLM that can basically talk, talk to you about anything. It can, you can take the conversation and direction. You can, you can, it's very general, so it can solve all those different tasks, but you interact with it through a very rigid interface. And so to me, it's just a matter of time before the interface itself becomes learnable and becomes, you know, part of the, the, the whole loop of like, you're not just delivering, you're delivering an application end to end.

51:15And that means you're delivering the language model, but you're also delivering the, the render and the pixels. And, and, and that's also a learnable component and the concept of applications might not necessarily, I think we'll need to figure out new abstractions for software. Um, the, the concept of application comes from this idea that you need, you know, separate kind of code basis to describe, to, uh, for, to power each individual, um, tool and each individual application. But you might, uh, you might think of something a lot more unified if you're, if you have a video model that's actually generating the interface as you go.

51:52Um, so it can take context from an LLM and allow you to combine kind of different, uh, different functionalities that traditional would live in different applications. So, so it's a, it's a way to solve, you know, software end to end, uh, effectively. We also see this as a powerful way to train computer use agents as well. Um, so this is, you know, one way to see this as, and in general with world models, there is those two directions. One is world models for humans and world models for, for agents to train agents.

52:25And so for every new work of, uh, world models that we do, we have this, these both uses become possible. So this is a powerful synthetic data generator for training computer use models. It could become, uh, a live, uh, RL environment that you could use to do online RL with, uh, with a computer use agent. Uh, and you can get wide diversity of different interactions, kinds of interfaces, uh, just generate on the fly that, um, to improve the, how robust the, your, your, your agent, uh, becomes.

52:56So that's the same also with the world models that we're working on for a robotics use case as well. Is there a research breakthrough that you're waiting for that would unlock the next set of use cases that you really want to pursue? Long context is a very important one. So being able to maintain consistency for long periods of time, and that depends on the use case. So for our characters model, for example, or for the interface world model, it's easier to maintain long sessions of interaction. If you go into more open ended worlds that you navigate and you take arbitrary actions in, we like, there is more, the context that which you can, and duration, which you can generate becomes limited much more quickly.

53:39So we see more degradation and error accumulation happening. Um, so the biggest challenge with autoregressive models is error accumulation is basically you're fitting generative frames back into the model to generate the next, the next frames. And if there is any small errors, they accumulate over time. That's not a new problem. It's a problem that LMs also have. And we've seen the, you know, the ability to generate now really, really long outputs. So it's a solid problem, but it's definitely still a slow challenge. Yeah. And what is the state of the art? Uh, so for, for Grok, it would be like 10 to 20 seconds.

54:12Of context going in there for video. With our characters models, we're able to generate up to 30 minutes of, uh, video autoregressively. Yeah. But that's just for the avatars. Yeah. So if we look at, um, GWM worlds, which is more our open ended world exploration model, um, it's, it's on the order of a few minutes, which is, yeah. It's probably enough for people because you have to cut to the next scene anyway, right? Yeah. It's not, it's not the ideal game experience if you have to restart every few minutes.

54:44So I think, but, uh, I think it's, yeah, for certain kinds of game experiences, you can, you can work around it. Uh, ideally you were able to just generate forever and it doesn't, it doesn't degrade. And I think that's a matter of time before we get there. Yeah. Genie has like one max one minute, you know? Yeah. This was your, you did a study on robotics.

Applying world models to robotics

55:05I think I also have just your runway robotics page though. Is this better? So last year we released Gen 4.5. So that was our latest base model. And we've been, as I mentioned, we've been doing all this work and world models and, uh, which essentially a lot of our approach to world models is how do you take a bi-directional diffusion model and make it autoregressive and make it accept actions. So instead of being a video you watch, it becomes a, it's a simulation that you step in and you can, you know, control it every step of the way.

55:37You can explore counterfactuals, like what happens if I take this action versus if I take this action and GWM one was the, it's the world model that we built on top of Gen 4.5. So we did all this autoregressive and like, uh, uh, distillation, uh, autoregressive and then step distillation on top of Gen 4.5. And one of the biggest use case that we saw for GWM one was in robotics. One thing we like to say is we, as we scaled video models, we accidentally, uh, created a state of the art model for robotics, um, by just scaling video models.

56:12So we realized at some point, um, meeting last year that robotics labs that are coming up to us and kind of asking to use video models for synthetic data, kind of asking us to post-train our video models to work really well. Uh, for robotics, uh, so that they can, they can use that to basically generate variations. That was the first use case that we saw. And then increasingly became clear that the models will be useful beyond just creating synthetic data to train robotic policies. They would also be very useful as simulators.

56:43So that means that you can use, uh, a video model online to test how your robotic action model performs. So you can take an action role and then get the outcome of the action inside the world model and then, and then continue that loop or like this close loop simulation. And you can use that to evaluate how well your robotics model works. Um, and the biggest thing that I think you need to solve if you want to build a simulator is establishing real world correlation. That if you take an action inside the world model, if you take the same action in the, in the real world, you get a similar outcome.

57:19So that was, that was the goal of some work that we did earlier this year. So if you go to the first link, so that was, uh, essentially we wanted to establish that, that, you know, uh, real, real to sim correlation for our world model. So that if you do a series of actions inside the world model, and if you do the same actions in the real world, you get similar outcomes. And we took our GWM one model and we, we use some benchmark data that there is this Roborina, uh, benchmark that's very commonly used to evaluate how well do different action models perform.

57:52And we use the same scenarios and settings and embodiments inside our world model. And we measure the correlation of how well did the action model perform inside the world model versus in the real world. And we saw that we could get very good correlation between, uh, between our world model and reality. And that means that if you want to evaluate how well your robotic policies perform, you can scale that much faster inside simulation instead of, instead of having to do that with actual physical hardware.

58:22And so that was a first indication that, that our models could be, uh, quite useful in robotics. And we saw as we're working with robotics labs, that that became like the first use case where they could use video models in a way that fit into their, their, their kind of training pipeline. Can I ask what the difference was from 4.5 to solving that? So the sim, the real gap has always been the issue, right? You train a robotics model on video data. It doesn't generalize to real world and the simulation had an issue.

58:53So it seems like you solved it, but how? Yeah. So a big problem with simulators is, you know, if, if you're trying to simulate rigid objects, like it works quite well. If you know, if you, if you, if you can describe the physics of objects very accurately, um, uh, then you're able to use, um, Isaac sim or, or Mujok or one of the traditional simulators, but for more complex interactions with, with a cloth, for example, or, um, like slippery surfaces, um, you know, with the, all the complexity that you, you want to be able to solve with a manipulation, with an action model that solves manipulation tasks.

59:30It's very difficult and so time consuming to build, you know, for each of those environments and each of those tasks, build the simulated version of that, uh, the digital domain of that, of that environment. Whereas with a world model, you, you just need to provide the first frame and then you just can roll out the poles inside the first frame. So, whereas, you know, we compared it to methods that required like 3d scanning and environment, and then 3d scanning each individual object before you can now, uh, you know, you can bring the dissimulation.

1:00:04Um, whereas with a world model, you just take a picture of the, of the environment and then, and then you're able to test how, how your policy performs. Our general thesis on robotics is, you know, there is companies that are leveraging a lot of teleoperation data to, to train robotics action models. There is now companies that are using, um, Yumi data, which is, uh, essentially human, uh, egocentric video where humans use kind of, uh, robotic creepers to perform different manipulation tasks.

1:00:37And then there's companies that are focusing on egocentric data, which is, you know, you strap a GoPro on, on someone's head and then you kind of capture them performing a task. We think that, you know, and all those are great sources of data for training robotics models, but the most plentiful source of video data is third person video data. It's an, if, how, how do we, as humans learn how to perform different tasks? A lot of it is by observing others perform those tasks. We don't learn from first person. We obviously do some trial and error and like, uh, to learn different things, but ultimately a lot of what we learn how to do in the world, we learn by watching other people do it.

1:01:16And that's how, when you're pre-training a video model, you're essentially doing that. It's a lot of third person video, uh, footage of people performing different tasks in the world, people doing sports, people doing household tasks. And our main thesis is that video pre-training, once you do that, you can then adapt a model to be useful in robotics use cases with lay way fewer hours of actual robotic data. So you require way less teleoperation data, which is very difficult to scale. Uh, and even if you look at egocentric data, which is a bit more easy to scale compared to teleoperation data, which requires actual hardware, it's still three hours of magnitude less of that, that exists in the world compared to third person video data out there.

1:02:01And so our thesis is, and generally like the most plentiful source of data will ultimately, ultimately wins their person video data pre-training is the right starting point for, uh, for models that, you know, you want them to generalize and be able to deal with new environments, new tasks, things that you haven't seen during training. That's the motivation for why we think our models are especially useful in robotics, uh, settings. And we've seen that to be the case, um, as well.

1:02:32You said pre-training there. So maybe it's like third person pre-training first person SFT. Is that like a curriculum that you can sort of introduce? Exactly. So if we look at GWM worlds, uh, so GWM robotics, so GWM robotics, it starts from Gen 4.5. It's the same video diffusion backbone, right? Exactly. Yeah. So you start from the base video model, the one you're using to generate, uh, cats and dogs and other interesting stuff. And then you, uh, fine tune on a very small number of hours of robotic data.

1:03:06So it's something on the order of kind of hundreds of hours compared to if you were to pre-train robotics model, the current pre-trainings go up to, you know, a hundred thousands or like millions of hours of, of data. And you're able to get quite, quite good performance, uh, quickly, uh, because the model leverages all the things that it has learned about the world and physics and human dynamics and the tasks that people care about from pre-training. And ultimately you, you want those models to generalize.

1:03:37You don't want to just be able to perform the tasks that they have during training and the diversity of actions and environments that you, that you have with a pre-training video dataset is much larger than, uh, what you can kind of realistically capture manually. How's the scale looking like for the post-training? Like you still want to do, is it like roughly 90% of the compute in regular video diffusion model and then scale up a lot or do it? Like we want different robotic models for different tasks or just the one base, really good world model can also apply to robotics.

1:04:13So currently, uh, we are post-training our models for specific, uh, embodiments that we, uh, for particular partners. So if they have a particular kind of single arm robot or bimanual robot, or shimanoid robot, we would post-train our GWM robotics model on their particular dataset. Over time, we see the different variants of GWM unifying, like I would expect, you know, if a year from now or two years from now, you have a single world model that can, uh, that can simulate manipulation tasks, it can simulate navigation, which is a lot of the gaming world models are navigational world models.

1:04:51You're moving around the space and it will also simulate kind of human behavior. So that's the character models. So instead of having three different models, you, you have a single model that's able to, you know, ideally you're able to simulate what it's like to be in the world. You're moving around an environment, you're maybe performing different tasks, uh, you're talking to other people. And that happens with the same, uh, a single kind of real-time video model that's generating that. Do you think you can solve self-driving? So if you are learning to drive a car in a simulator, you have a world model.

1:05:24Your robot is basically, you know, car can manipulate so many axes. How far off are you from something like that? Or a really good ADAS system? World models are definitely being applied to, uh, self-driving, uh, research. Right now, mainly for evaluation use cases, but our focus has been more on robotic manipulation. We've done some work on AV, uh, world models as well. Uh, but yeah, we do think that world models are, uh, video models are the best starting point for both simulators and also policy and the action models.

1:05:58So that's, that's the other side to this is that once you have a great world model, then you can just add an action head and it can predict actions as well. One way to think about it is if you take the starting frame of a scene with a robotic arm and you ask, you know, you prompt the model, generate the arm picking up an object. It would, and if it generates an accurate enough video, then it should also be able to generate the exact poses, uh, in 3D that the arm should take to, um, to perform the same action.

1:06:30So this is the direction that's now the, the popular term for it is world action models, which is you're starting from a video model and then you're adding an action, uh, action head to predict the actions and it becomes a policy essentially. One thing I'm also impressed by is how much data you actually need to train these kinds of models. You probably can't say exactly how much, but like, uh, you know, like the, the original, uh, diffusion models. And from what I know, even of the open source Chinese models, it's not that much data.

1:07:00Isn't it surprising? What do you define as much data? Um, uh, yeah, it just comes, goes in there. Is it token count still relevant? So it's a bit more complicated and, uh, What is, I mean, just gigabytes, right? Yeah. Hours of video. Yeah. Yeah. Yeah. I feel like something that's interesting is, it seems like the, let's call it tokens to param counts in language models has really, you know, maybe they're three years ahead or whatever. It seems to be a lot higher than, uh, video models still, even though technically video has more information, you know, per, per bit.

1:07:35I don't know if it seems intuitive or maybe there's just a lot of, uh, like the, the variability between a pixel to the next pixel is not that high. So like maybe there's just a lot of information that is repeated. My answer would be, it's still very early. Like, like the training video models will scale way further than it currently is. And, and you'll have capabilities that go much further than the current models can do. So one thought experiment that, uh, I like to use, it's, it's almost like the Turing test of video models are like the Turing test of world models.

1:08:08Uh, um, uh, I call it the lucid dream test is you have, uh, I mean the actual person lucid, lucid dream, it comes from lucid, lucid dreams, right? Lucid dreams is telling your dreaming while you're in a, yeah, exactly. So lucid dreaming is when you realize you're inside a dream and then you basically play around, be able to control what happens. No, there was also an inference guy called lucid dreams. Yeah. So let's say you have a VR headset and, uh, you, and, uh, you're in a room with, uh, and you're wearing a VR headset and that VR headset, you know, most, uh, two days VR headsets have a pass-through mode.

1:08:47So you can see directly what's in front of you in the world, or you can obviously render kind of something inside, inside the VR headset. And there's going to be a point where those interactive real-time video models become good enough where you wear the headset and you're in the same room and you're walking around and you're kind of, and you're interacting with objects. You're able to kind of move freely in that room and do, um, and interact with any object. And at the end, someone asks you, did you, were you using pass-through mode or were you actually, or was this, uh, rendered or generated, uh, footage?

1:09:24And if you cannot tell for sure, if that was what you were seeing as you were interacting and moving around the world was generated, or it was, uh, kind of pass-through mode and it was just what was happening in front of you, that's an indication that the models have become good enough. And we're not, we're not close to that yet.

The path to general simulation

1:09:41And a lot of it is just this idea of really simulating dynamics and counterfactuals. Well, like, you know, if you ask a video model to generate a person scoring a goal versus a person failing to score a goal, it would do a better job at scoring the goal because there is a bias from the training distribution. There's a lot more videos of the person succeeding at scoring the goal, but if you have an interactive model, you want it to be able to generate counterfactuals. Like if I take this action versus this action, you want it to generate equally realistic outcomes.

1:10:14Uh, so that's, I think the big gap between video models and, and, and world models is that idea of the counterfactual generation. And if you want a great model for robotics, you want to simulate failure very well, because whether you're using it for evaluation or you're using it as a, in an online RL loop in the future, you want to be able to have the model kind of try and fail to do things and, and improve. Um, and so in order to do that, you need to be able to simulate things failing. This is the only domain where you have too many successful examples and not enough bad examples.

1:10:48It should be easy to generate failure. Oddly enough, I think like early image video models weren't good at being human realistic, right? Like you see a lot of the high res 4k, like professional photography, but not just everyday life, like normal picture, right? Everything looks like it's professionally generated, like professional pictures, but not just like normal, like, you know, messy cables on a desk. Okay. So there's, there's this stuff, uh, you know, one thing we also covered that you guys have, uh, video agents that you launched, I guess, how does the traditional, let's call it frontier L like, you know, auto regressive LLMs,

1:11:27like feed in, you know, to, to all this, they're driving, uh, your robotics models or they're driving others, your video agents, uh, production, um, any, anything where you, you see the overlap of auto regressive and diffusion, let's call it. Yeah. So harnesses are really important across all those different use cases. So we have this video agent, which is essentially an LLM that is very effective at tool use of different, uh, image models, video models, and kind of helps you through creating a, a project end to end. So, you know, very often in like a traditional kind of advertising flow, you have a brief, you start from a, and then you generate some, a storyboard and then you generate the video, a video agent, and a runway agent kind of helps you through that whole process.

1:12:10And it helps you also analyze performance data. For example, how well did this ad performance of this ad, and then generate me more of the, based on those learnings, figure out what, what, what to generate. We think that the harness is a very important piece of the, the pipeline, uh, as I mentioned, all the video production, uh, all the, all the production video models use some prompt completion that happens. And we expect, you know, that to become more and more complex and more, you know, you generate longer and more detailed descriptions before you use the, the, the diffusion transformer.

1:12:44I do think eventually, you know, there's increasingly this unification into Omni models where you have the, you're training the models end to end to both do autoregressive text prediction and also, uh, diffusion as well. So you're predicting the next token, uh, of like, you, you, maybe using some reasoning and planning of the scene, and then you're passing it into the diffusion head that's actually generating, uh, generating the pixels. Yeah. I think it's currently maybe only Gemini and Quen do it.

1:13:15I'm not sure who, which of the Chinese models are Omni, but yeah, it's, it's not, it's not a very well, um, popularized modality, I guess. It's an interesting use case when you think about it, right? Because not only do you have to end at like language model reason, diffusion head generate, you don't have to output there. You can go back into feed that output to the same model, reason again on improvements, and it can do a lot of loops just in its own. Uh, I guess the question is like, do we need that or can we just do agent scaffold, like do it outside the model?

1:13:52Is there a big benefit to doing it in? I think there's generally the trend of something is first done by a harness and then it becomes part of the model, right? So you had a chain of thought prompting where you had to do this super detailed system prompts to get back and output. Yeah, let's think step by step. And now the model basically generates the reasoning trace by itself before it gives you an answer. And in the, in video models, similarly, a lot of the video models of kind of the early days were single shot video models. And you had to use some kind of orchestrator to turn, you know, generate multiple shots in parallel, um, and then turn it into an actual video.

1:14:28Um, and comfy UI and just all over the, you know, all these nodes. Yeah, like a spaghetti, uh, workflow. So, um, and now you have multi-shot video generation where you have the, you directly generate multiple shots. And there is a benefit to that because then the video model learns some, you know, to generate a single shot wall, you need obviously to figure out a lot of stuff about the world. Uh, to generate multi-shot video wall, you also need to basically get some like video editing instincts. Like you need to figure out what is the right pacing of shots.

1:15:00And also LLMs are not that good at it. Like they're not that great video editors. If you ask, uh, LLM to take some videos and then kind of auto create, uh, edited video out of that, uh, it would, it would feel uncanny. So I don't think LLMs are actually that good yet at being video editors. And I think there's benefit to learning that end to end. Um, so I would expect, you know, the training general is the things that, you know, you need a harness for eventually get kind of injected into the, into the model itself.

1:15:31And you learn that end to end. Do you find that you need to hire engineers who can, or researchers who are also artists to infuse that taste? Or do you have kind of artists in residence to distill them? Um, we have, we have a, uh, a, a large creative team that's very actively involved in the, in training those models, um, like on the, you know, in every part of the way. And like, how do you caption videos well? So that you capture the stuff that you need for like the cinematography, the aesthetics, the camera direction in as detailed ways as possible.

1:16:04So that you're able at inference time to actually, uh, at least see that through the model, you know, we have a creative team also does a lot of evaluation of like, you know, what constitutes a usable video out of those models. And so they're very involved kind of through every part of the process. And I think that's one of the special things of runway is just that, that mix between like kind of creatives and researchers kind of sitting by side by side and kind of working together to, to build the next generation of our models. I think that's been a really, really important piece to, to, um, you know, how we've operated as a company.

1:16:37Yeah. In some senses, well, you can only do this in New York. I mean, you have other offices, but like, you know, I'm trying to find some poetic, uh, significance. In the fact that you are a big New York company. As there's a few parts of being New York, obviously there is that intersection of all those different industries and like media kind of advertising, uh, like. Yeah. This is very advertising. The art, like the, um, the art scene is New York. Not, not to, not to say anything bad about San Francisco, but it's, uh, there is more, more going on.

1:17:08There is that component. There's also, I think we benefit from being outsiders and thinking of things a bit differently. Like not being in the same, like hive mind of, uh, ASI, uh, of, uh, of Bay Area and like taking, and also taking our time to, you know, to get where we are today. Like building the, growing the team intentionally and bringing people who are, yeah, both on the creative side and also an engineering research side. There's obviously a huge talent pool of amazing people in New York.

1:17:38So that, that hasn't really been, been a problem. I mean, congrats on everything. Uh, what are you hiring for? You know, what should people look forward to, uh, for the future of Runway? We're hiring across the board. Uh, I think this is probably the most open roles we've ever had in the history of Runway. Uh, we're growing our research team quite significantly. So if you're, um, if you're excited about video models, if you're excited about world models, if you're excited, especially about robotics, the robotics team, we're hiring roles in the robotics across, uh, kind of software, hardware and research.

1:18:10Um, so definitely, definitely reach out. And, and, you know, a lot of people don't have direct robotics background, but what should they have, you know, if they want to be useful in robotics? So ideally some experience with, uh, learned policies, uh, would be good for, for robotics. Uh, but we, we tend to hire generalists as a, as a philosophy and like, uh, people who learn really, really quickly. Uh, but some experience in the income of domain expertise in robotics is something that we're, we're definitely looking for, uh, for the, for the next months.

1:18:45Um, and then we're scaling the, the go-to market team significantly. There is, um, a wide, like very, very active enterprise adoption happening around video models at the moment. And, uh, we're really trying to, uh, respond to all the demand. Yeah. Great. You want to talk about the, uh, open source robotics stuff? Sure. It was just random notes we had. Uh, NVIDIA launched Cosmo. I guess it's interesting. So, uh, you're a founding member AI labs to build open source world models in physical AI.

1:19:17Um, anything much to talk on here is open research. The biggest thing is that, as I mentioned, world models are still, uh, early, like there's still so much that you can scale and those models further so much more advancements and, and things that we can figure out or how, how to improve those models further. And I think this is, uh, it's important that some of this research happens in the open and figuring out what is some incentives for different companies to come together to, to actually bring some of that research into, into the open and open source.

1:19:48And so Cosmos coalition was, uh, initiative that we co-founded with NVIDIA to bring some of that research as open source. And that could mean open weight model releases. It could mean benchmarks that measure physics and things that people care about when building world models. It could mean infrastructure. So really how do we grow the ecosystem of world models and make that something that also it's easier for a developer, a researcher that's just starting out, that is excited about world models to kind of contribute to the field.

1:20:18I think there's some amount of like, is this also our response against the Chinese world models that are being released, you know, or, uh, is there not part of the consideration? I do think it's, it's important for, in video models, you know, if you look at the, the leaderboards of video models, I would say right now, the majority of models at the, you know, the top 10, top 20 are Chinese models. There is, uh, only a handful of companies that are made into the leaderboard from like the U S or the West.

1:20:50We're doing better with images, but with video, we're very behind, right? And so I think, I think it's definitely important that we invest more broadly as a community to make sure that we, we can, those models can, you know, we, we have competitive models, uh, out there. Well, like what was the stop us from distilling from them? I don't know if that's the best, uh, long-term approach that you're bounded by the performance that you can, it's, it's almost a bit of a pessimistic, uh, view that you can get better. You know, you can, it's free data. Like it's, you know, it might as well, like if they're doing it for a lot, for the text language side, they might as well do it for the video side the other way.

1:21:25Yeah. I mean, I do think we're, we're quite capable of training great models without, without the solution at the moment. Yeah. Yeah. So anything you have to say on benchmarks and evals, like I feel like what I'm hearing is a lot of people really like arenas for video and image models, uh, customers and whatnot as well. They only want the best on the leaderboard and they refer to arenas a lot more than language models seem to do, but any, any notes on benchmarks? What's lacking? How does the average person compare? Well, these both look really hyper-realistic more than that outside of, we did talk about like robotic simulation, um, the physics and all that, but anything to say?

1:22:04I actually think it's the opposite in some ways. I think people generally creatives and artists and marketers, like people that are using our platforms, I think rely less on, uh, arena scores. And it's, it's just so easy to, you know, generate with a bunch of different models and then compare the results visually. Like one nice thing about image and video models is you can immediately tell with your eyes, like what, what feels good from an aesthetic standpoint, like any artifacts and issues with the physics of those models, you can immediately tell.

1:22:35Um, and so that's, it's actually easier, I would say to evaluate, uh, as a human, uh, there is also those models and it is in, in language models where you have those, those very complex kind of math and coding and, uh, tests where it becomes a lot more harder. I think for, for humans to evaluate and kind of discriminate between the, the performance of frontier models at a time. So I think in practice, people just test out the same prompt with a bunch of different models and see what the results look like.

1:23:06And right now in runway, you can use our models and you can use third party models as well. So it's, it's very easy to do that.

Creatives and the future of tools

1:23:12Amazing.

Creatives and the future of tools

1:23:13We're going to end with the AI, Runway AI Summit. The last sort of social societal issue, I guess, I don't know if, if this is a, uh, thing is the, uh, you are at the tension between, sort of artists, uh, creatives and AI, a lot of people in that, that community hate AI. Obviously the people that are in the runway community don't mind using tools. Uh, it's just another brush, but, um, how have you seen the sentiment change? I mean, our perspective, yes, it's just another branch, uh, brush is just another camera. It's, you know, it's the latest of a long generation of, of tools.

1:23:46Technology in art and art and technology have kind of evolved together. I think there's been a pretty significant shift over the past few months and it came, uh, some of it, you can see with a lot of public figures speaking out in favor of, of AI and being, you know, like in, in Cannes, you saw a few, a few directors speaking in favor of AI. We had, uh, Ron Howard in our film festival, there is, uh, Mark Soscheze also adopting AI models. So you have more of those stories coming out every day of like, uh, a well-known figure, um, kind of speaking in favor of AI.

1:24:22And it's just a matter of, in, in my mind, it's, those models are becoming more and more demystified. I would say I have, I have also a, a bit of a hot take that one of the things that made the, uh, the initial response to those models, maybe a bit more heated than it needed to be was this idea of text to video of, you know, you have a single text description and you get back a, a full, a full video. Um, yeah, there, there was a misconception, obviously you can generate it to our feature line film, but the models of today now take a lot of references.

1:24:55They take, they are very controllable. And I think when people see, uh, a tool that allows, affords many degrees of freedom and control, they respond to it differently. And it matters less that it's a generative model than, uh, than the fact that you can actually steer it to the direction that you want. Um, and so I think when people look at, you know, complex workflow on top, on top of those models, when they look at, you know, all the ways in which you can steer them.

1:25:25And you can provide now with some of the latest models up to 50 references, like the conversation becomes a bit different because it feels much more like a storyboard, a tool versus like something that a magical kind of entity that figures out like the, your entire film for you. Any notes on, like workflows changing for people in the field? Like, I think engineering at least has had a lot of people where they're like, expectations have changed. I'm, you know, 10x, 100x more productive and you can get a lot more done.

1:25:57Uh, same thing is, you know, you're making dev tools for creatives. Um, any notes there? Like there's some people that don't want to adopt, some that do like anything? Yeah, so I think in terms of like what people care about, uh, I see that we have gone through a few stages. So we started from a stage where the main thing that people were looking for was quality. Like, uh, as you know, we scale those models and the quality improved dramatically. That's something that people still care about, but it's, it's now in addition to controllability, like being able to steer those models with references, with, um, you know, different kinds of inputs and storyboards.

1:26:33And now my sense is increasingly people are going to care about latency more and more. As those models become better, the ability to iterate very quickly becomes more important. And like, if you can, you know, with a single prompt generate 10 different, uh, outputs, like almost instantly, you can explore way faster than before. And you get some of the magic that characterized the, you know, the creative tools of the past, like Photoshop was instant. Um, and, um, and with, we love some of that with generative models, you, you're waiting for two minutes to get back a video.

1:27:04And I think we're going to bring some of that back now with, uh, with the, the real time models. Yeah. Exciting. Exciting. Uh, the last thing we'll plug is this one, uh, runway AI summit. You're finally doing this in SF. Uh, yeah, so, uh, we're very excited about this. So this is, uh, in late September, September 30th, we're doing a summit on, uh, primarily focused on physical AI and real-time video generation. We have panelists from Nvidia, physical intelligence, the BotCo, DeepMind. Yeah. It's going to be, I think a very interesting series of conversations.

1:27:37We'll, we'll try to make the, the panels really technical and, uh, elicit actual kind of substantive discussion and hopefully some, some interesting kind of disagreements and interesting debates on, on things. And, uh, yeah, the, there's tickets available. Hope, hope people can join. Since you mentioned it, uh, what kind of disagreements and debates should people think about? What do you expect? So it's things like, you know, the, there is, uh, one debate right now in the robotics world is, uh, VLAs versus world action models.

1:28:11Okay. Um, so there is labs that are really, really betting on one of those two directions. Um, there is a, what is the best source of data to train robotics models? There's just the third party, first party that we've talked about. Yeah. There is, you know, the people who really believe in kind of further scaling tele-op data versus leveraging more large scale video data. So those are kind of some of the, and then there is, you know, the, the world models debates of predict pixels directly versus something like JEPA versus a more 3D based, uh, 3D based approach.

1:28:44Um, so I think we're at a nice time in world models because there is still that kind of active debate happening or like, what is the best long-term direction? And I feel very strongly that it's video predict pixels directly and, and scaling video generation models is the right approach. But it's, I think there is a lot of interesting, uh, debate happening, uh, by, by researchers on like, what is the best, best, best, uh, kind of path to take. Like, it's interesting that it's all on like, sort of, let's call it the policy layer and the data model layer is the, the physical side is completely solved.

1:29:17Like all the sensors, all the actuators, all these things there, we, we have everything that we need. I don't think that's a self either. I think it's definitely, um, different problems. You know, it's like, I want to dream of all these things and then I get, uh, you know, I, I buy a robot or I, but I try to assemble my own and I can't even get the, you know, the, the, the motors to like work right. Right. Like, and it's, you're dealing with very sensitive, uh, equipment that has, uh, voltage and power and like heat and all these things, which, you know, abstracted the way we're sitting here, we're talking about software and talking about models, but like really you have to deal with those kinds of things too.

1:29:56Yeah. And, uh, I think, um, I'm, I'm generally also not, um, not opposed to incorporating other modalities into our models. Like we've seen, the simplest case is they can generate video and audio at the same time. So they can generate RGB and they can also generate, uh, they can generate sound and audio, but my, you know, I've written about this as like, what, what does the maximalist version of a world model look like is you're incorporating more and more modalities from the universe and you're training a model.

1:30:26on different scales of observation as well. So, yeah. We've got a good, uh, good essay that people should be doing real world. No, uh, Meta released a model that was like six modalities in one, right? I forget what the name of the thing was, but it was like, yeah, okay. Death is one of them, but death is like a transformation of RGB in some sense. Image bind. Image bind, yeah. What other modalities? They had heat. Audio, depth, heat, text, uh. Whatever IMU is. Um, I do think like you, you might as well do ultraviolet. You might as well do like whatever other modality you feel like, cause it's all data to the model.

1:31:01Yeah. And a, a big bet is also that there is transfer between all those modalities. So one of my favorite, uh, examples, which is quite old at this, at this point is there was this fine tune of, uh, stable diffusion. That was called refusion. Yes. It was the music one. Yeah, just fine tuning, uh, stable diffusion on spectrograms and became a quite capable music generator, right? So there is probably some spatial patterns or like spatial temporal patterns, if we're talking about video, that kind of emerged at different scales and different modalities. And so there is some degree of, you know, meta learning that the model has done that, that allows you to learn faster.

1:31:37If you, if you start from a, just a model train on images and train it to predict audio, than if you train from scratch on just, just audio. Um, and there is some other interesting examples. So there is this, this project called the well it's, uh, it's a data set of physics, numerical simulations in physics and biology and a bunch of other domains. So it's, so it's essentially different physical systems across very different scales of space and time from like astrophysics to low level kind of, um, like atomistic interactions.

1:32:11And we've seen, uh, we, we've done some, some, some, some, some work on this and, and we've seen that we can take our video model where, you know, real world video looks nothing like this. And you can actually fine tune it on, on those numerical simulations and just treat them as RGB frames. And you actually get reasonable performance much quicker than if you just train from scratch. I think, I think we've seen this across languages. Yeah. Deep SQL CR as well. Yeah. Deep SQL CR. Like you don't have to tokenize text.

1:32:42Like you can just throw them in as images. There's a lot that happens in that base pre-training. Like there was an argument a long time ago of people saying, oh, humans have so many, uh, sensory representations, right? Smell, touch models have a whole, two more modalities that will never, like that we don't even have data for. And it's like, okay, you take AQI sensor, like you can, you can try this stuff, but actually, you know, there's so much happening in just the base train run that you don't get as much from these little things. Yeah, exactly. And I think that's what it solves is data scarcity. So you don't have as much, you have so much video data available, but you don't have, you know, like olfactory data.

1:33:21The cool thing is it goes the other way too, right? So if you want to do physics, like if you want to measure this or you want to have a diffusion model do audio, it transfers really well. So like in your case, the little bit of post-training for robotics gets a video model to use its fundamentals in another domain. So we can apply that to other stuff too. Yeah. And if we look at, you know, like how do you make those models more useful for in scientific domains? And if, you know, if you look at AlphaFold, it had all this very, because of the data, you know, the limited amount of data that it needed to be trained on, it's basically, it's very fine-tuned architecture just to solve a kind of protein structure prediction.

1:34:02But if you take all those disparate sources of scientific data and you bring them together under a single model, like I think that's an approach that can help us kind of solve new kinds of problems across science by leveraging all the learnings from one modality or one set of data to another. So very early days for that direction. But I do think that that's where ultimately what the end game of kind of simulating the world is. You're not just using RGB, you're using RGB as a starting point, but you can incorporate more and more modalities of the universe and leverage the transfer that happens from learning from one to the other.

1:34:41I guess the follow-up there is what's the drawback of Omni? Like why, why is everything not an Omni model? So why not now? And why would you start from language backbone or image video backbone and then go Omni from there? Does it matter? Yeah, we need to take it one step. We need to solve robotics first and then we can go into... Solve everything now.

1:35:04Yeah, I mean, I do think there is a lot of open-ended research that needs to happen for those Omni models. There is, you know, there is a lot of things that require careful consideration when you're bringing multiple modalities into a single model to predict. But I think, you know, I expect those to be solvable. Wonderful. You've been very generous with your time. Congrats on all your success. And I'm excited for the AI Summit or physical AI Summit. Yeah, thanks for having me. And people should check out the film festival if it's in town, right?

1:35:36Yeah. You'll be going to be touring all over the place. Yeah, next year we're probably going to do that. So we do film festivals every May or June. And we did the last one in New York, L.A., Tokyo and at the AI Engineer Fair. Yeah, yeah, yeah. So, yeah, hopefully even more places next year. No, I think someday, you know, you will be hosting the Oscars of AI video. And, you know, I think people should take this very seriously as like a potential career they can have. The Oscars of AI video will be called the Oscars.

1:36:10All right.

1:36:12All right. Thank you. Thank you. Thank you.