Emad Mostaque, Chief Executive Officer, Stability.Ai
Interviewer: Jeremy Kahn, FORTUNE

Brainstorm A.I. 2022 - Generative A.I.: Silicon Valley’s Newest Gold Rush

Conferences

Generative A.I.: Silicon Valley’s Newest Gold Rush

Welcome back. I hope you all have a chance to stretch your legs and grab a coffee. And now let's continue with our main stage conversations. There's probably no hotter topic in ai these days than generative ai. It's an umbrella term for software that instead of just analyzing existing data can create original blog posts, Wikipedia entries, artistic images, video code and more up next a demo of the remarkable tech from stability ai followed by a conversation with the firm's founder, Ahmad MMA Stock. He'll lay out the economic and creative upside that will follow from making this open source free to use code available to anyone who wants it. Please welcome to the brainstorm. Ai stage. Amada stock founder and ceo of stability ai we'll get to the demo in just a minute. But first of all I want to ask you, why did you do this and especially why did you do this as an open source project? I mean potentially you know, this could make a lot of money and film entertainment industry, why do it open source for fun and profit? I think I saw this technology and this technology classical AI was built for one thing there that can do anything and I realized it would be like servers and databases which are almost all open source and it's all about scale and service customization. So I thought rather than wait five years to get there, let's just go there at once and allow it to spread everywhere, which it has. So like for example, stable diffusion has just overtaken Bitcoin and ethereum on get hub stars in terms of developer popularity, The fastest growing open source software ever. With the biggest community building. The craziest things, wow. Um now one of the things when stability sort of 1.0 came out, there was a little bit controversial is that unlike some of your competitors like dolly, you didn't apply any filters to either the training data or the output to make sure that people couldn't generate pornography or other sort of not safe for work images. Now with the two point oh version that was just debuted last week, I believe. You know, you've actually rolled that back a bit. You have applied some filters that make it much harder to generate pornography. Why the reversal on that. So the first part was actually a research initiative led by the LMU. This team at the University of Munich with collaboration from ourselves runway Ml and some others thus providing the main computer infrastructure and Robin Rombach, our lead generative code are doing it and they said, let's start it as a base and make it just an image of the internet and see what we can do. And that's what came out. But we put a safety filter on it and an ethical use license as well to limit the amount of output for our own release of stability. We decided and listen to the feedback that we would filter out not safe for work elements. So more people could use it. We decided that we would also make it a more open data set. So we can really understand how the model worked. We spent millions of dollars kind of doing that as well, which is now results in the model that anyone can use in a commercial and corporate setting. Great, well let's let's let's see the model. And this is a special demonstration. This is the first time I think you've shown this particular version of the model. Can you, can you explain what you're gonna be showing us? Sure. So basically, um we released 2.12 days ago, released the original one august 23rd release 2.1 tomorrow. Um this is the place of kind of innovation here to put one is very special because when we released 1.1 point four, actually in august 23rd, it was 5.6 seconds per generation. We released that produced that 2.9 seconds for an image. Um just a few days ago, as of next week, it will be 30 images in a second. So we designed it to be fast and now we've got to real time. This unfortunate isn't quite real time, but we've got the one second generation and again it can generate anything. So you take words, you put it in and for example, is on stage between yes, you can do a lovely labrador, right labradoodle. We like doodles. Labradoodle and then in the second it generates a picture of multiple, lovely labra doodles with a hat. And then boom, it generates it with a hat. I think it's just loading on the internet and then we can do by van Gogh and it becomes by van Gogh. Um so this is actually a very interesting thing because 100,000 gigabytes of image is compressed into a 1.6 gigabyte file that works on your Macbook. And as of last week it works natively. So you'll be able to generate these images in a second in a stained glass window. So you can generate just about anything because it understands two billion different concepts. And this is just unimaginable. Right? And like that. As of next week, it will be 30 frames a 2nd, 2.1 adds back a lot of the artistic elements because some lovely stained glass windows. Isn't that nice? But then it's also very flexible. So for example, what? We can do it the wrong way, We can do, let's do nine images here. We took four images of jeremy and we decided to create a jeremy can model this is just with four images. Who knows if this will work? Uh a picture of jeremy Kahn as an astronaut. You wanna go to space and then we can do HD matte painting, why not live with the new technology. Who knows if it will work or not? We'll find out in a second. The suspense is killing me how we doing there. Maybe we should make a professional actually it wasn't professional enough. It's interesting because it understands all these kind of modifiers. If you do a professional not professional, you want to negative prompts. You can say not ugly and it becomes not ugly. Um You gotta do not ugly for me, definitely jeremy like any of these like how is this one here? Yes, not bad, not bad. There we go. Just four images. We've got jeremy there and now we obviously do by Van Gogh. Right, so I think you know, super powerful technology that can create any image unimaginable. Just that's not that's not me. That's not another jeremy's emerge. Yeah, so matt professional, let's do photo with some 35 millimeters studio lighting to make it good. So I mean I think this is the thing, this technology is still nascent, right? This is unimaginable. Just a few months ago, let alone in the last year, it's just evolving at such a rapid rate. Still not used. Well, it's closer but lovely. This technology is designed to be kind of zero shot where you just put something in you get out. It's part of a whole design process. So we have people seen people take this and create entire movies from it because they take frame by frame like into the Spiderverse and they transform it into an anime movie. We've seen people take this and create architectural designs at scale and you know, again it's part of the process. It's amazingly powerful tiny file. We put it out to the world. And now people are developing around it, developing ecosystem that they can then take and build really useful applications. That's what you need actually useful things. This is why the number one app store, app one app store of the year is one bow which uses our technology. Number one on the app store right now is lens, which creates these types of customized profiles based on this version of the technology as well. So I think it's gonna be very interesting to see how this develops going forward. Absolutely. And there, I mean there are skeptics out here. I would say this is great. It's a lot of pretty pictures. But what's, you know, what's the real business impact of this? So like if you look at movies for example, this will cut down movie generation time by half. So we have an example of creative director was shooting a movie in the desert about serial killer with some very famous actors and actresses, serial killer had a book of all the victims that he had and they're gonna have to hire 25 actresses to fill that book, put them in the seventies makeup and you know, fly them out. Sag rates make up all this kind of thing. $600,000. He trained a custom model overnight on seventies pictures. It was done, save $600,000. That's just one example when you look at video games, movies, other things that transforms them any type of design process of working with automotive manufacturers to create brand new car designs. Because you can do these kind of dynamically, you know, you can just type in like an awesome super car in SAn Francisco and it will generate those but you can customize it with your own assets. This thing when you take the standard model and customize it, it becomes super duper powerful. That's an awesome super car in SAn Francisco. And are you thinking of using this beyond just images? Are you thinking about using the same underlying technology for generating other sorts of things? Yeah. So we have the most popular language model in the world. DPJ Neo and others from our Luther ai kind of community and we actually do models of every single type. So we have audio models that can generate any type of music. We've got three D models coming video models. Ultimately the whole thing is that there's no barrier to creation when you have this next generation of models from a big data age. We took big data extrapolated for a specific models. These very general models to understand principles like understands the word supercar SAn Francisco. We talked about jeremy maybe it needs more than four images. We have jeremy, right? And this is the next stage in AI and the next stage in real usability of this technology and for you guys the stability, what's your kind of business model going to be going forward, Every single creative industry in the world is gonna have to react to this because like I said, as of next week, this is 30 images a second, which is basically video brand new experience where you can create anything dynamically. We're going to go to them, take their exabytes, petabytes of data and convert them into models because again, 100,000 gigabytes to 1.6 gigabytes. Something quite crazy. Yeah, absolutely. Um I said I wasn't gonna take questions in the audience but I think I will because we have a little bit of time. Who has questions for Ahmad? Anyone have questions right down here in front, wait for the mic to get to you. Please please stand and stay here you are. What would you say to you can you can you are my name is Ian you tilly attention life. What would you say to creatives because this can either augment a creative development of what they do or they can kind of feel like I know that you can take my job. So on the optimistic purse, how are you encouraging creative and creative teams to leverage this technology to do what they do better. So I think this thing, what is creativity? Right? So this can match together concepts. So we still have an iterative process to get there. There's just one part of the process. So the ideation phase, the fine tuning phase, the kind of like we have text, image, image to text in painting so I can remove, I can make that car blue and it'll make that car blew in the next version. You forget how to integrate this into your process is just as you did with after effects Photoshop, a lot of these other tools because it's just a tool, it doesn't do anything by itself. So what we found is the best artists got even better and this is what we saw with and there's an Ai that beat at least a dollar go who is the best go player? The average go score, it looks like this over time. Is that because they all got better because they use the technology as a tool to extend and embrace it because it's here and you should use it right now and then you'll outperform all of your competitors. Right. Other questions, we're here. Hi there, Ahmad Rishi Garg, a partner at the Mayfield Fund. Um, uh, you've got a lot of other projects on the website that you're working on, you've obviously started with Image generation. But one of the things that's captivated everyone's mind is text text threading and text generation because of the broad applications. Could you talk a little bit about the future horizon for stability and all the different applications you want to build into and how and how kind of starting with image Get you to a place where you can do multimodal applications over time. So that's excellent as a business perspective when I saw the image I saw as we're going from a 10% to 90% like if I show you this a year ago, it's impossible, right? And the fact that it's real time now is also impossible. So you wake up and do some impossible things before breakfast text. By contrast we got 85% to 95% there and I struggle to see where the edge is going to be when you have a chat bots that can do anything like use chat. GPT somewhere, you know it's fantastic. All the Ai models will be like that. Whereas an image I saw that there are giant asset stores of I. P. That need to be converted because you always want mickey mouse or halo or any of these other things so we can create everything, something becomes valuable but it's not enough just to have image cause image will be standardized instead. What we're doing is we're building the ready Player one Oasis slash Holiday where anyone can create anything they can imagine. And that's not only media, no more power points, no more slide creation in a few years. Just describe it and make it more peppy, make it nicer and they'll do it automatically recreate game of Thrones season eight because it was a bit rubbish. You know these types of things is what we are enabled seamlessly. And you talk a little bit about how you build stability itself because I think a lot of people are surprised you guys almost came out of nowhere with this fantastic technology. I think there are a lot of businesses out there where they, you know, they talk all the time about the struggle they have, hiring ai talent, you've obviously been able to build, you know, really cutting edge stuff. Um how did you do it? So we built a community when some of the large open source models were open source claiming the big entities could do it and I looked at whereas the arbitrage properties and the gap, the gap was super compute data and talent. The talent came from our communities and we hired primarily from our communities. The 1st 96 employees, only one PhD So differentiated talent then data. We built large open source datasets like Lion, 5.6 billion images. The previous large, 100 million to power this. And then finally, Supercomputer worked closely with amazon to build the 10th fastest public Supercomputer in the world. And so that was overwhelming amounts relative to what happened. So 4100 fast in the UK was 6 40. And that unlocked all of this because we allowed people to experiment and play and figure out and make mistakes that then optimized for this crazy model right here, but also other models that were released in our generation to the new language, ones to the others. Because you need all the components all in one and that's why I created stability to be like this foundation layer for next generation generative AI. And you you now have a sort of a partnership with AWS, is that right? Yes. So this has been a long term partner now, it's accelerating up, we're gonna power generative AI through kind of amazon and make it available to everyone. It becomes a sort of differentiate for different peoples clouds. Obviously has got their relationship with Open AI AWS as you guys. It's kind of interesting. Exactly. You know, and there's different approaches like quite integrated, very close to Microsoft where we're just trying to get this out to the entire world, right? Thank you so much for joining us. This has been fantastic. Really exciting stuff.