Cris Valenzuela, Runway CEO: A Fortt Knox Update

Yeah. Well, first of all, thank you for having me. It's great catching up. Yeah, you know, so we've been working with the Nvidia team for quite some time, really closely on both research and testing of new models and new chips. And with Rubin, we were I think one of the first ever to port a video model to it. So we took 4.5, Gen 4.5, which is our latest video model we released only a couple weeks ago and it happens to be the best, still the best video model in the world. And so we took it and we were able to port it to this new both architecture and platform of Nvidia's Rubin in just one single day which is kind of a big deal to be able to take what we already have and immediately accelerate it with like the newest and the frontier kind of chip Nvidia has been working on with great performance which is exactly what we need and want to see.

I mean it's hard to tell because there's a lot of software kind of things behind the scenes, but training a model is a complex thing to do and porting it to architectures like even new chips is relatively complex. We're talking about months of work and having backward compatibility where you can take our existing architecture and model that we train and running on Nvidia's Hopper which is like the previous generation and porting to Rubin I think in a day kind of shows that it's not about months or even weeks of work. It's just very seamless work and hopefully you can do it without too much work in the future as well which happens to be the case for us. So I think we're saving at least a couple weeks of work because it's backward compatible.

We're still running the tests, but we're seeing significant improvements in speed of inference. We're testing mostly inference at the time. Of course, it's also going to work with training, but for now, we're only doing inference. And we've seen really good updates and really good performance improvements on text to video or image to video, which is the ability of a model to generate sequences of video coherently for 10 seconds or longer. Videos are particularly very memory intense and compute intense tasks much more than language models are. And so, you need a chip that can handle both the memory and the compute conditions and necessities of a model like ours. And we were able to see some of the same performance with less of the hardware that we used to use before. So, which is great.

Yeah, I mean look, we're pretty focused on building the best video models out there and building the best video models require us to train them in multimodal ways. So Gen 4.5 in rankings right now if you compare them to any other model out there from the biggest research labs to labs outside of the US, Runway still ranks the best. And I think part of it is we're really focused on making sure that the model works really well on both video and audio together, but also aesthetically, efficiently and performance-wise. It adheres to the prompts the users are putting in it. It's fast at rendering the videos in a good way. And also the cost is really good. We've seen so far that and this is some of the collaborations we've been discussing with Nvidia is in order for the models to become really accessible they need to be really good at the unit economics of how people are using them behind the scenes and so that requires us to train them in such a way they become really effective and efficient at being used at large scale for our customers and that's something we obsess with. So this is the trade-off between high performance but also at a low cost. Having both I think is the sweet spot and I think 4.5 and the future updates that we have are continued on that trend. And world models which we released on the heels of 4.5, we also released our first public world model, also builds on all of the learnings and training we've done with our video models and so we're also now porting our general world models into Nvidia's new architectures to gain and benefit from those performance improvements.

I mean right now you have every single media company and advertising using Runway and this platform for both advertising, creating end to end ads, for Hollywood and media, creating scenes or much more complex shots, to entertainment at large. I think that's the lowest hanging fruit and the most obvious use case is there and everyone is using it. Like I mean everyone is, trust me, every single media company has an AI strategy or has used an AI model or used Runway in some way and it's going to only grow from here. I think for us it's like well how do you also scale that to others, like we're now selling to architects, we're selling to gaming companies, we're selling to everyone who wants to make pixels on a screen. Robotics for us is an interesting new growth opportunity where you can take these models that have understood the physics and how the world works and now you start deploying them to train robots and so you're making, this is kind of interesting, you're making videos not for humans to watch but for robots to watch and learn from. And so really for us there's a lot of opportunities here and I think world models at large are going to be perhaps an exterior of AI starting now.