コードのいらないアプリ？ Runway「Solaris」が描くインターフェース世界モデル

Solaris takes a different approach and it just renders the pixels directly instead of requiring some code that renders the pixels. It has learned how software should behave as a result of interaction. And so it directly takes mouse and keyboard input and then it produces the pixels. So you can start from a UI design that you made in Figma and immediately animate it and be able to interact with it. So for designers that want to very quickly prototype, like what does this interaction feel like, it's a much faster path to getting there. Like, long-term persistence is definitely a challenge, and that's something we see especially for the more open-world experiences that we build. And then the other challenge is latency, just making sure that you can get the interaction latency to be as low as possible, because it can easily break the experience. Welcome, humans, to the Neuron AI Explained. I'm Corey. He's Grant. Hi, Grant. Hello. Hello, hello. Welcome, humans. Well, we're going to have a cool chat today with a company you know on a topic that we rarely discuss, world models.

But before we dive in, today's episode is sponsored by Dell Technologies and NVIDIA. You'll hear more about them here in just a bit. Yeah, and to have this conversation about world models, we are talking today with Anastasis Terminatus, co-founder and co-CEO of Runway, about Solaris, their new interface world model that generates interfaces frame by frame instead of rendering them from code, and the much bigger bet underneath it, that world models can simulate software, interactive worlds, robots, and eventually even more of the real world. Wow. Yeah. Runway helps push generative video from an interesting research problem into a real creative product category. And now the company is arguing that many of the same models can become interactive simulators. From websites and apps to game worlds, which I'm really excited about, video agents, and robotics. So today we're going to unpack what's real today, you know, maybe where some of the rough edges still are and where it's all going to go next. Please take just a quick second to like today's video and subscribe to the channel so you never miss an interview or one of our live streams.

And on that note, Anastasis, welcome to the Neuron. Good to be here. Thanks for having me, guys. Thanks so much. We're glad to have you. So you've said that Runway isn't a creative tools company suddenly switching paths to world models, that this is really an extension of the work you've been doing for years. And I'm wondering kind of what's the through line from the earliest runway models like Gen 4.5, GWM, and now through to Solaris? Absolutely. So just to give some context, so at this point, we've been doing research in video generation for more than half a decade. So it's been a long journey of just starting from those models generating very low resolution, very low quality distorted outputs to actually now generating videos. And we've run user studies on this. People very often cannot tell apart from real footage. So it's been a long journey of essentially scaling those models

and improving the architecture and the data and a lot of research across the board to get us to this point. I think from the very beginning of building video models, we thought, like, ultimately, what does it mean to predict video? It means in order to predict video well of, you know, a human doing a task, you need to understand so many aspects of the world. You need to understand physics, and you need to understand how objects interact. And that understanding can be useful beyond just generating videos and towards building interactive systems that allow you to test different scenarios. And so we've seen over the past couple of years, we've been especially focused on this direction of world models, which we see as an extension of video models. And it's essentially interactive simulators of the world, that are based on the research that we've been doing in video. And we think the same framework applies to robotics,

and it's very useful there, but it also applies to building new kinds of interactive experiences in the photos models for humans to use. And the work that we've been doing in Solaris is one example of that. So world model is one of those terms that a lot of people have maybe perhaps conflicting definitions or they get a little bit confused about it. It can be very broad. What would you say is like what capability makes Solaris a world model in your definition? And how do you define it and kind of separate all the different definitions of world models? So ultimately, the way we define world models is they're learnable simulators of environment. so they allow you to test what is the outcome of specific actions you take in that environment. And the environment could be the real world, but it could also be the digital world that's happening on our screens. But as long as it's interactive, real time, and takes actions as the input, like continuous actions,

that essentially needs a definition of a world model. So actions can mean a lot of different things. In a robotic setting, it could mean like kind of moving the robot arm and kind of, you know, changing the state of the gripper to grab an object. In a digital environment like Solaris, it means the action is a mouse click or like an interaction with your keyboard that you're making. And so Solaris is essentially a simulator of all the software that we have on our screens. it's trained on a large scale data set of screen recordings and all sorts of interactions with software and it essentially learns how software should behave as a result of human action without needing to generate code to render the interface That's so wild So like in a traditional app

if I click a button and some amount of deterministic code changes some underlying state. In Solaris, what holds that state? And, like, how does it know what should happen next? Yeah, so essentially, so the traditional workflow of software is, and especially front-end development is you write some HTML or CSS, and then that then gets rendered into pixels. or nowadays you use a coding agent to write the HTML or CSS for you. But Solaris takes a different approach and it just renders the pixels directly instead of requiring some code that renders the pixels. So it has essentially learned from data, from just looking at many, many, many recordings of interactions with different kinds of software, with websites, with games. it has learned how software should behave as a result of interaction.

And so it directly takes mouse and keyboard input as input, actions as input, and then it produces the pixels. And it kind of, if you click on a button, the model has learned from just pixels how the buttons should behave, what should be the result of that interaction. If you, you know, if you have like a, And similarly, kind of with the game world's, you know, models, it's kind of similar where if you have an initial frame of a character in an environment, then it knows how the models should behave as a result of like an arrow key navigation. So I want to move forward. It knows how to predict what happens next. So here's the thing about enterprise AI right now. Every leadership team has the pilot. Every company has the proof of concept. But actually, scaling AI from a developer's desktop to the data center, all the way out to the edge, that's where things start to break. So the hard part is getting all of those pieces to work together.

The data has to be ready. The infrastructure has to support real workloads. The agents need guardrails, and teams need to know where each AI workload should actually run. That's why we teamed up with Dell Technologies and NVIDIA to build a resource hub we're actually excited about. The hub gets into the problems teams run into once AI leaves the lab. How do you get enterprise data ready for RAG and AI agents? When does it make sense to run agents locally on desk side systems? What changes when you stop treating AI as a pilot and start treating it like a production infrastructure? It also covers the architecture side, why the AI stack needs to work as one system, how to avoid moving data more than you have to, and why open-weight models are becoming part of the enterprise AI conversation. And the whole thing is backed by Dell AI Factory with NVIDIA, a full-stack approach that brings together AI-ready infrastructure, software services, data platforms, and NVIDIA acceleration. So teams can start small, scale across the enterprise, and build AI systems that are governed, useful, and ready for production.

So if you're tired of hype and just want to understand what it actually takes to make Gen AI work, head over to the Enterprise Guide to Scalable AI hub on techrepublic.com, or you click the link in the description of this video. Now back to our interview. So does it ever retranslate that back into any form of deterministic state or code, or is it always just keeping that, like, the closest thing I can think of not understanding it fully is that it keeps it in its model weight, like it maintains it there. Like, does it regenerate when it looks back the way it's already looked or something? Yeah, or does it ever make, like, a carbon copy of it so it can go back and reference it? Like, it's very interesting to think about. Yeah, so it maintains a history of frames as context, so it knows what the previous interactions that the user has already made with the interface. But it's also, we see that Solaris can be the front end, and Agent, NLLM, can be the back end. So when you're making a specific action with Solaris

There is a VLM that's running in parallel That's basically interpreting that action as well And so if you need to maintain some state Like you can have a markdown file where you write Like the user did X You can definitely do that And so we see those being kind of used in combination But unlike traditional interactions with LMs which I think we're in this paradigm where the LLM is super flexible, like when you're talking to a JavaScript or Claude, but the interface at which you're interacting with the LLM is very rigid, and it's just all very basically the paradigm. Your HTML, CSS page. Exactly, yeah. And so why shouldn't the interface and the visual UI, why shouldn't it be as flexible as the underlying LLM? And that's kind of the idea of Solaris. Very cool. Yeah, so you're basically taking code out of it. Like you don't need code in this world, basically.

Yeah, it's all end-to-end. It has been the lesson of machine learning. Like you remove all the intermediate steps and just learn the mapping from input to output. And the input is the users interacting with a mouse or keyboard. and output is whatever gets back shown on the screen to them. Totally. So if I was to be using an interface and I was to ask, could I just ask it, hey, can you change this? Can it look different? Will the VLM and the world model combination be able to change it for me? Yeah, so there's a lot of interesting things that you can do with this framework. So one is, you know, we have this idea of model behavior in LMs, right, which is I want the LM to be, you know, to be very concise or to write me in a more formal or casual tone. And this idea of model behavior translates also into interfaces.

Like some people prefer, you know, very bold visual design. Some people prefer, like, very information-dense interfaces. and you can have just this same concept kind of translated into Solaris where you can essentially system prompt what kind of interfaces that you want to interact with and customize them based on what your preferences are. Is there a future where a person would have those preferences stored in a world model and surf the internet and things be catered to them in the way that works for them? or is that a ridiculous question? It's okay if it is. It's a great question. I mean, you know, now we have, like, LMs have become incredibly good at tool use and the whole ecosystem of MCPs. And so if you have an LM as the backend,

you can basically use Solaris to generate an interface that applies into all those different services and tools. and so this basically becomes the visual interactive front end to your LM or your agent and it could be for some use cases text and the chat interface is fine but a lot of use cases that require a lot of visual exploration or learning about new things you benefit a lot from a rich kind of visual output So if you want to learn about any concept in biology and you want to visualize the internals of the cell, or you want to book a trip and you want to visualize the city and see the different attractions, there is basically a ton of use cases where having an interface that's generated on the fly is very useful. But it can plug into all the existing ecosystem that we have for LMS that's growing every day.

Wow So basically if this was to happen The internet could essentially just turn into Correct me if I'm wrong A bunch of markdown files and API calls Essentially And maybe databases somewhere in there Databases of markdown files And then the rest could just be generated by this Yeah, I mean I think it's a very plausible feature And it's I think both for more professional work interfaces but also for entertainment and for gaming. I think there is just so much that can be done with world models to change how we experience games and interactive entertainment. And so it provides essentially a unified framework that you can think of all kinds of software can become kind of generated by a world model. What specifically, like, if you could give an example of the kind of thing that Solaris can do that a traditionally coded project cannot

in terms of, like, what does all of this extra, like, real estate essentially from not being chained to the code, what does that buy you in terms of speed, what you can do? I think so one initial thing is that it allows you to iterate on ideas for interface design much quicker. So you can start from a UI design that you made in Figma and immediately animate it and be able to interact with it. So for designers that want to very quickly prototype, like what does this interaction feel like, it's a much faster path to getting there. and we've done a few user studies where we essentially compare, we have an LL, we give an LL a visual design and say kind of recreate that in code and it can do a good job at recreating the static design but the moment you interact with it, you immediately like basically start to see the flaws

versus like with a video model that has learned motion, really good kind of motion interaction and like how software should behave. we get much higher preference for the outputs of Solaris. So from that perspective, just like prototyping and very quickly testing how different interfaces feel like, that's something that's very, you know, it speeds up the process quite significantly. And then for customizing the experience of a piece of software to your preferences as a viewer, I think it gives you a lot more flexibility there as well. And then I think it's just all those use cases I mentioned, like any use case that involves visual exploration. So if you want to build an e-commerce storefront, you essentially can simulate, you get much closer to the experience of actually being on the physical store than you can with a more traditional interface.

We've been working with a lot of e-commerce customers for some early design partnerships where essentially you can build virtual try-on experiences where you're really pulling clothes from the virtual rack, and you can very quickly visualize them. So you can build a lot of really natural, interesting interactions. and then responsive design is another big piece of just with a video model you can adapt to different aspect ratios a lot more easily so that's also what we're excited about, there is all sorts of screens in the world beyond your laptop or your phone, there is screens on your car or screens on smart devices and having and instead of having to build like specialized layouts for all of them,

with a video model you can adapt to the parameters of the screen much more easily. That opens up a totally out there question for me that I've been thinking about this whole time, which is at what point does this combination that we're talking about of like world model plus visual model break us free from screens, right? So right now we kind of get attached to screens because, you know, we're used to those being either where we process information or our viewports. But I'm imagining a world where maybe at some point, if a lot of this computing is done in the cloud, then we have holograms and or, you know, projectors, you know, projectors like what we have today, that then makes the entire, you know, 3D world around us a space that could potentially be used as a screen. It's holiday. Yeah, basically. I mean, how realistic, unrealistic is that, that this kind of paradigm, if it goes this way,

could lead to some version of that where, you know, five years from now, we're not talking to each other on a laptop, but we're talking to each other, you know, from a projector and anywhere in this space. And the world model is just generating that. On a fake golf course or something. Yeah, wherever. Yeah, so we haven't yet run our world model on a holographic. What are you waiting for? I do think there is, you know, like some of the bottlenecks to getting wider adoption of mixed reality, virtual reality has been software. And it's just like it's much more difficult to build really great experiences for those devices compared to kind of your 2D screen. And so I do think that world models have a big potential there. Of course, it's, you know, it's a year of VR every year for the past 10 years. Yeah, right. Don't hold your breath. I can explore it than I expected, and so I do think there's a huge potential on kind of our boring 2D flat screens.

There is a huge potential for wall models there. But I am very excited about how those models can be used in VR, and it's something that we do want to explore in the near future. That's awesome. Yeah. Even I was watching the GPT-6 Astra video, and, you know, they're just talking to GPT-6 on a projector. And, you know, you could just imagine instead of having a computer, you just have a projector that you'd watch a movie on, but you do most of your coding there. It makes me wonder, you know, what other interaction paradigms could we add to, you know, if we follow this thread and we're using world models to generate the interface, you know, could we add, like, voice as a modality? Could we add interaction where, you know, now you're not relying on a mouse, but you're actually clicking with your hands and gesturing. Yeah, I do think voice is a very natural input for a model like Solaris. And also one other thing to note is that our model can natively generate video and audio.

And so there is a lot of interesting use cases where you can incorporate audio and sound effects as part of the interaction. And then we have a separate real-time model called Characters, which is real-time video interactive avatars that you can talk to. And I think in the near future, we hope to merge those two quite a bit. You can have an interactive avatar that helps guide you through an interface that's generated by Solaris. So I do think there's a lot of potential of voice as the modality that you interact with. Wow. What's the hardest problem in front of you on the world model side of things? I mean, is it long-term memory, determinism, compute, back-end integration? Like, what are the biggest challenges these face?

Yeah, so, like, long-term persistence is definitely a challenge, just making sure. And that's something we see especially for the more open world experiences that we build. I would say it's a bit easier on a model like Solaris because you're a bit more constrained on the kinds of interactions you can have with a UI versus if you're navigating an open world. Like a video game world or something, yeah. Exactly. And so making sure that if you enter a room or if you leave a room and come back to the same room, that the details of the environment are maintained. Those kinds of challenges are very important to solve. And it's just challenges around kind of handling very long context. In my mind, they're pretty solvable. We made a lot of progress in being able to generate kind of consistent, long rollouts. But it's definitely still a challenge. And then the other challenge is latency, just making sure that you can get kind of the interaction latency to be as low as possible because it can easily break the experience to, you know, if you even like a half a second delay, it's unacceptable for an interactive interface.

And so making that work well is very important. And I do think eventually the model that will work the best is some hybrid approach where there is a lightweight model that kind of renders at very high FPS that runs locally and then a cloud model that kind of is more powerful and kind of generates some more sparse context that's provided to the lightweight, very fast model that's running on device. So there is definitely some hybrids that are going to become relevant in the future to get that latency down as much as possible. So you don't realize. Yeah, I was even thinking, like, you know, you could imagine a version of this where you do still use code to some extent. You know, maybe it's just a midterm solution where, like, you can have the renderer agent happening in the foreground. And then, you know, when you leave a space, you have, like, a really fast coding agent on the back end, like, sketching it in. so like it's like kind of filling in you know even if it's just like for a video game for example

like the placement of where the objects are and labeling them or something like that like almost as if you were making a mock-up of the game originally before you do all the rendering on top of it you could totally have the coding agent kind of like put all those like shapes in place and stuff so that it would however it would work and then use the renderer on the front end very interesting there lots of interesting combinations you could do when you combining different things together like this Yeah totally There's definitely a lot of interesting combinations within LLMs and video models, and that's what we've seen with our traditional video models. A lot of people use them in conjunction to LLMs to kind of help them write great prompts or kind of help them brainstorm through concepts. But in a real-time setting, that becomes very relevant as well. So, for example, with our game world model, you have LM on the background that's basically generating the dialogue for the characters in the scene

or generating different events that should happen next. So there's definitely a lot that can be done that interplay between LLMs and real-time video models. On your Solaris evaluation, I noticed you compared generated interfaces alongside frontier LLM code-based standard interfaces, and users preferred Solaris on instruction following. I think it was natural behavior. I was wondering kind of what does that benchmark tell us, and maybe what doesn't it about, what did you learn from that? I think the biggest learning was kind of to our prior discussion that LLMs can be very powerful at reconstructing static interfaces, but when it comes to reconstructing motion and kind of really getting a fine-grained interaction right,

I think they're still very hit and miss. So the biggest gap between LLAMS and video models right now is that understanding of physics, understanding of motion. That's where we see that video models are still very powerful. Even if you look at some of the demos of Astra today, You have these really incredible demos of just using MS Paint to reconstruct some photograph of a person. But if you look at some of the demos that's been done with Astra to reconstruct the motion from a video, it's much more rough. So there's that thing with the Apple parkour video where you have Astra trying to recreate the motion of that. And it's impressive, but it's definitely not there.

Yeah, or just trying to get it to animate characters in video games is the hardest thing to do. You can make pretty good models in Blender now with something like the GPT-6 Astra or even the new Claude models. but try to get it to animate it to look like a normal person or walk normally. It's super goofy. So that's definitely not the ideal use case for LLMs. Exactly, yeah. And the same finding applies to robotics as well. So there is a lot of impressive demos over the past couple weeks of Astra performing some pick-and-place robotics tasks and doing quite well. but we've seen that for more fine-grained manipulation tasks, there's still a lot of benefit from leveraging a video model. Nice. Yeah, let's actually talk about the robotics work. So you recently reported a 0.95 correlation between simulated and real-world rankings across eight robot policies, but you were careful to say that this isn't the same thing

as perfectly predicting absolute success rates. So what evidence would convince you that a learned world model can replace a meaningful share of physical robot testing in your experience doing this? And perhaps you could just share a little bit more about what you're doing in that space. Yeah, so what we saw, and that happened about a year ago or a bit more than a year ago, we started getting a lot of requests from our robotics labs to use our video models as part of their workflow on either generating synthetic data or for evaluation. So just to give some context, on robotics, the biggest challenge by far is data collection. In order to scale data collection, you need to buy more hardware. It's very time-consuming. It's very costly, and it can be very unsafe as well. And very often the best way to collect data is to actually deploy those robots in the real world, but in order to deploy those robots in the real world,

they need to be good enough. So you have this chicken and egg problem where you're not able to get the models to be good enough to deploy so that you can gather more data. And so this is exactly what video models are great at. It's just generating different scenarios from the real world. And so another example is in AV, generating most driving footage that's out there is very boring. nothing exciting is happening there's no crazy weather conditions there's basically all these very long tail scenarios that rarely ever happen but are very critical for any model that's deployed in a self-driving setting to be able to handle well but for a video model it's very easily able to generate what is the same driving footage look if it's snowing or if there is heavy rain or if there is kind of a pedestrian kind of going in front of the car.

Like all those difficult scenarios, it's able to generate quite easily. And so we see video models being very critical for advancing robotics and just physically more broadly. And we've been working over the past year with robotics labs to help them both on generating data but also evaluating how well robotics models perform, which is also very time-consuming, just testing whether your model works well and handles all kinds of best cases. So that particular study, we essentially took... So there is this very famous benchmark in robotics called RoboArena. It's essentially the equivalent of LM Arena for robotics, and it tests different robotics models on a variety of manipulation tasks. So it's things like folding clothes or pick-and-place tasks or writing on a blackboard, things like that.

And basically we measure the correlation of running a control model in the real world that already has happened in RoboArena versus running the same model inside our world model. So evaluating how well a robotics model performs inside kind of the imagination of GWM1. And we saw that if the model succeeded in the real world, it was very likely to succeed in simulation and vice versa. If it failed in real world, it was very likely to fail in simulation. So the overall outcome is that you can essentially evaluate robotics models much more cost-effectively and at a much larger scale by doing that inside the world model instead of doing that in the real world. Okay. Something I've wondered for a while is we don't get a lot of hands-on time around traditional world models

or in your case, this first interface world model, which is really awesome. Do you think that it will follow a similar trend to what text models and image models have in terms of pricing that makes it increasingly accessible over time? Or do you think this end of the technology is probably going to remain more expensive than others? So I actually think it's going to be cheaper. So running real-time models is cheaper than running a Frontier video model. You get the quality that you can get from a real-time model compared to our Gen 4.5, which is our kind of flagship video model, is not as good. It's maybe more specialized at the moment of either interface model or a game model or a character model. But because we have to run them in real time,

we need to optimize them quite heavily and we're able to serve them at a much lower price than we serve the traditional video models. And that's essentially, if you're running it, if you're using something Solaris, it's a continuous interaction that happens over minutes. If you were charging for that, the same way you charge for getting back a video from a video model, then it would become very expensive very quickly. So we think it's one of those cases where it's cheaper, but it's used for a much longer period of time. It'll be cheaper, but you'll be using a lot of it, one of those kind of like basic tokenomics right now, that every time an efficiency gain is found, It's like demand quadruples overnight and just completely negates all of the efficiency fines. And that makes sense to where you'd get to a point to where like an individual unit of your intelligence will be like a bearable cost.

But you're going to want a lot of units. Exactly. If you look at interactive entertainment, for example, there is a lot of use cases where you want basically content that's generated on the fly based on what you want to see, or you can take even existing pieces of content and I want to view it from a different angle, or everything becomes interactive with real-time models. So there is all these kinds of experiences that I guess will be built on top of those models that we're especially excited about. I want to put a couple of ideas together to try and see how all this fits. So, you know, we talked a little bit about the worlds. I don't know if we mentioned GWM Worlds 2 lets you define a world, its subjects and its rules, and then you can steer it with freeform actions while it's generating the video and audio live, kind of like what you were just talking about. Now, do you see the end state for, you know, let's call it,

we've talked a bit about how this could work in, like, let's say, like replacing the Internet as we know it today. But for something like robotics, for example, could the end state there be one foundation model where the action space changes depending on whether you're controlling a game or whether it's, like, actually onboard the robot or it's being streamed into the robot? or will all of these domains always need significant specialization, like kind of like what you're talking about now, where you have the world model, you have the character model, you have the robot model. And building on that a little bit, if there is something on board of the robot, kind of like human brains, will there be multiple models all stitched together? And it's not like one thing. Because there was a study that came out last week that found out that humans actually have two brains after all, not just one. So I'm wondering, maybe we just need to create AGI or whatever comes from here by combining a lot of these different models, kind of like we talked about with the Internet idea,

where you've got the video model and the world model, or sorry, the VLM and the world model. What's your take on that? So even if you look at GWM, we do have those variants that are post for specific applications but they all derived from the same foundation model And we think ultimately it going to be a single model that serves all those use cases And there's a lot of benefit and transfer between getting good at video games versus getting good at interacting with the physical world. And so we do think that a unified single model is the approach that works best. this was just what allowed us to serve today most efficiently to specialize models. And then to your second question, right now there's definitely some benefits in robotics in particular where you have almost like system one and system two where you have kind of high-level planning

which decides, you know, if I want to cook a meal, there is the, you know, the high-level steps of the recipe, and then there's the fine-grained actions of, like, cutting the onion and, like, doing all these different actions as part of the process. And those, right now, are best served by two different models. One is more a VLM planner, let's say, and the other is the, you know, the world model that basically knows how to, is very good at those fine-grained motion planning. But, again, machine learning is a story of doing things more end-to-end, and Unity is kind of the name of the game. So I do expect that that's going to continue to be the trend if you have a single model that's actually able to do both. Cut out the middle model. Yeah, cut out the middle model. so runway still has this huge creative product surface with with gen 4.5 for generation i left

2.0 for editing and you've got act two for performance capture and characters for real-time video agents how much research transfer is now flowing in every direction between all of these different tools along with your world model work? Like it feels like you have cool parts that could feed one another in a lot of ways. Yeah. So it's on the research side is really all very unified. Again, like 90% of the capital expenditure of computer runway goes into building the base model and the base model gets on post-trained to be a real-time interactive world model. So it's really the same underlying foundation model. And we think that pre-training is very critical and it's where the model learns physics, where it learns most of the capabilities it needs to have to be a great world model. And so that's also what we see with other labs that are just taking an existing pre-trained video model and then trying to build world

models on top of it. It's a much more difficult path because a lot of the relevant things and the important things actually get learned in pre-training and post-training is mainly for eliciting those capabilities in a more consistent way or for optimizing it for something like a real-time deployment. And so creative tools and the work that we're doing on the existing product is growing really well. We're seeing a lot more adoption of AI video this year as just a result of those models becoming really good. And that's across the work that we're doing with the Hollywood studios. It's the work that brand teams are doing to basically accelerate their marketing and advertising. I mean, we see it every day. I think that the amount of AI video out there is increasing every day. And getting better to watch. I'm always watching, like, AI short films or, you know,

I saw a guy who had pushed a pilot just a couple days ago. He's been working on it for quite a while. that kind of stuff has come a really long way. And you've been involved with that. You've worked with some various studios and things along the way as well, haven't you? Yeah, so we work with all the major Hollywood studios. We work with most of the major ad agencies. And so this is, you know, we had one of the first of its kind partnership between an AI company and a film studio back in 2024 with Lionsgate. And so we've been very close to that world for a long time. And the perspective on AI has changed quite a bit, especially this year. I think for a long time, Runway was a tool that everyone was using, but nobody wanted to say they were using. Yes, it's true. People have become a lot more open about their use of AI. And I think it ultimately has become a lot more demystified.

Like, it's just a way to iterate faster, a way to kind of turn more ideas into final outputs. It's not, you know, vision still matters, taste still matters. There is great AI video out there and not so great AI video out there. So the same rules apply. It's just a new kind of camera tool. How long until we see a full feature-length, fully AI-generated video from a major studio without giving anything away, just your general prediction? I think I have too much insider knowledge, too. That's fair. That's fair, but that's kind of an answer, so it's interesting. I do think people ask me when for a long time Runway's North Star goal was to get to the point where you can

generate it to our feature-length film I remember mentioning that a few years ago at panels and people would just think that it was 10 years away we were so far off. And now it just feels, it doesn't feel like an ambitious goal anymore. What is next? And that's a lot of the North Star goals now is just like, how do we build even more capable simulators? And how do we really get to go beyond just generating realistic video towards being able to iterate in real time and rethink software, rethink how we can approach physical AI and all those new applications. Recreate the internet, fully solve robotics, and then also create the holodeck. These are my challenges for you.

Good luck. I think we'll still be running Kubernetes even in that far future. I believe it. I believe it. True. I think, and this is me personally with absolutely no insider knowledge in relation to this, that I would be shocked if there's not a major motion picture in a theater 12 months from now. Like, I feel like I will see my first movie in a theater in the next year. The question is whether or not they'll admit that they did it or not. Yeah, we're probably already seeing parts. Yeah, yeah. That's my hot take on that. I think I have one last question for you here, which is, unless you have something else, Corey. No, you're good. Three years from now, what would be the clearest sign that your world model thesis is actually right? Like, what do you want to see us doing differently with computers, media, or robotics that just isn't practical today?

Like, you know, in the near term, what do you think is likely? I know with predictions it's really difficult in the AI industry because it surpasses everyone's expectations again and again and again. It's not prediction friendly. Yeah, but if you were saying, like, okay, if we hit some of our North Star goals here, like, what do we actually want to see possible in three years? So I do think I would be very surprised if world models are not being used in production systems in robotics in three years, probably a lot faster than that. You know, across there is obviously a wide range of difficulty of different tasks in robotics, but there is definitely a lot of tasks where we think world models will be useful in the coming 12 months and we'll start seeing them deployed in production. And so that's one area where I would kind of bet that world models will become really useful in robotics very soon.

I think when it comes to interactive software, first of all, if the default way of using video models is not real time in a year or two, I would be also very surprised. It just feels, it's just such a much better interaction where you can immediately visualize what you, you know, what you want to create. And so real time should become the default of video generation. To me, that's obvious and it will lead to much wider adoption of those models because latency and cost is definitely a barrier right now. And then I do think with some of the work that we're doing in Solaris, if there is, you know, I would expect a large share of the software that we interact with every day to be kind of generated on the fly by a video model or a hybrid, kind of an omni-unified model. That's also, like, would be my expectation. And it might not be every single piece of software would benefit from this,

but for a significant kind of share of daily tasks, it just feels like it's going to allow for much richer, much more customized and personalized interactions. Hey there, humans. Quick heads up. The Neuron is going IRL in San Francisco on November 18th, starting at 4.30 p.m. If you're in town, come hang out with us. We'll have drinks, a live recording of the Neuron podcast, special guests, and a whole bunch of AI people together in one room. Huge thanks to Slack and Alumni Ventures for helping us bring the night to life. It's free, but space is limited. So scan the QR code on screen or hit the link in the description to RSVP and save your spot. We'll see you there. That's so sick. I think the whole operating system should run with that. Like, the GUI should just be replaced by a world model. I think that makes total sense to me. It's just willpower. People will make it happen. Anastasis, thank you so much for joining us. It's been a lot of fun. Thanks, guys.

What's the best way to check out what you all are doing over at Runway? Keep an eye out for more news from Solaris and all the many things you do. Yeah, runway.com is where we're at. We tend to announce things basically all the time on Twitter. Like we try to have a very fast cadence of releases. I noticed. We have a ton of open roles available across research, engineering, go-to-market. So if you're interested in seeing some of those advancements and get access to that insider knowledge, then please take a look at our careers page as well. Excellent. Very cool. Well, if you liked today's video, please take just a moment to like and subscribe. One more time, I'd like to thank Dell and NVIDIA for sponsoring today's episode. If your team is trying to move AI out of that pilot phase and into real production workflows, Check out the Enterprise Guide to ScalableAI at techrepublic.com.

It covers data, infrastructure, operational decisions that make enterprise AI work at scale. You can also click the link in the description below. Thanks so much for joining us, everyone. And that's all we have for today. So farewell for now, humans. Thank you.