Why AI Matters w/ Emad Mostaque | EP #52

0:00By next year, end of the year, I believe you'll have ChatGPT on your mobile phone. See that one more time? By the end of next year, you'll have ChatGPT on your mobile phone with our internet. Today, the world changes. This is a regime change. Four of the top ten App Store apps in December were based on stable diffusion. 41 % of all code on GitHub right now is AI -generated. We have figured out how to make human scale. So you have this amazing thing that can create anything. How far out are you able to see? Can you have a sense of predicting where we are in 10 years? I can't see plus, but there's enough fear and some of it is deserved, but there also needs to be hope.

0:37Humans are humans. Bring them the information that creates the most value and you will change the world.

0:46I'm excited to share with you one of the most extraordinary conversations I had at this year's annual abundance 360 summit. But it was with a friend someone I've had on this podcast before, Imad Moustak, who's the founder and CEO of Stability AI. You know his company from a number of his products, including stable diffusion. We sat down to talk about his moonshots to talk about how generative AI is changing the world. What do you do if you're a 20 -year -old just getting involved or if you're a seasoned CEO or entrepreneur? and how is this going to change the world of Hollywood, education, health care?

1:24You know, it was an incredible conversation. I'm excited for you to hear. You know, my mission through this podcast is to help you hear the voices that are filled with optimism and abundance. If that's of interest to you, then please subscribe to this podcast. All right, enjoy the episode. By the end of this episode, you should have a pretty good idea where you stand on AI. Are you excited about it, you're fearful of it. I'd love to know. Feel free to share your stance with me on Twitter. Hey pal. That's a good one buddy. Listen, I know you have so much going on right now. It's insane. We are in the midst of the most extraordinary period of technology ever and you're in the thick of it.

2:08Yeah, I think what's it? A good phrase is everything ever all at once. Yes. Yeah. I can't mentioned the conversations he's been having the last 30 minutes but it would blow your mind. I'm going to give you this clicker. You have some slides to show. I think the very first question to ask is if you were going to sort of bring people up to speed on where we are right now, would you share some of what your thoughts are here? We have figured out how to make human scale. How do you make human scale? That's the most difficult thing. Basically, classical AI was all about taking stuff and extrapolating it.

2:49We figured out how to figure out principles and latent knowledge with huge compression in these files, as we'll talk about in a bit. Like, the most valuable thing you can have is a good EA or a really good analyst. I have Esther, if she's amazing. Make such a difference. I'm sure all of you in this audience will realize that. We now have that. Basically, what we've done is we've train these things on large amounts, what huge amounts of data to understand principles. And now when you use a GPT -4, it's like a really talented intern analyst with a bad memory. And we're about to fix the memory. And it knows how to pass the medical bar exam or medical licensing and others.

3:25It knows how to draw beautiful pictures or create audio of any type. And it's basically available to everyone for pennies. And so we have to think, what does that mean? It can mean huge disruption, or it can mean basically we're at the point of utopia. and abundance. So if you don't mind, I'd like people know stable diffusion and you'll show them the chart that really rocked me when I first saw it. But what are the wide range of things that Stability AI is working on right now? Where do you want to go? Can you tell us about it please? No, I can tell you. Basically, my aim is to provide the building blocks for a society OS.

4:05Yes, every part of society, if you have skilled experts in them, our mission is to build the foundation to activate human and human potential. And across every single modality we're building models, so audio, video, protein folding, DNA, chemical reactions, language, and others, we're doing it for every sector, so you've got a bank or GPT, a board GPT to replace all your crap board members, you know, all the way down to national models. So we're doing an Indian national model at the moment and Indonesian one a Japanese one so that everyone can have a File that they put words in and images audio text anything pops out that's appropriate to you to enhance you So I want to I want to click on that one because it's very important people talk about Bias is in the system But what you're talking about is creating like what you're calling a foundational model That is biased to you or to your company or your country.

5:01Yes The existing internet has centralized artificial intelligence that guides our Attention our memorization where we go with this technology we can push internet intelligence to the edge So a single person country culture company can have their own a eyes And that's an amazing thing because it can work for you not against you And so that's the personization because what is bias right there are inherent biases in our society and things We can never move fast on this thing. We can't move fast enough to address them But then what if you could have your own stories? Everyone here has a story. You're members of A360.

5:35You believe in abundance. There's a whole variety of things that make us who we are. Have the AI respond to your own story and work for you. I think that's the intelligent internet and that's the amazing future we have coming. That is. You built Stability AI as an open platform. Which is very different from everybody else. I mean, it's interesting, right? OpenAI has that name. and very funny, if you go to openAA .com, it doesn't go to OpenAI, it goes to stable. Well, actually Open .AI here. Open .AI, sorry. Which is pretty funny. And appropriate, because you are the OpenAI platform. We are. I mean, actually OpenAI .com also works.

6:18But, yeah, I mean, basically where we were was that we have this explosion of like research. Look at this is, again, exponential, of the amount of ML research that occurred. And we were like, let's build a generalized intelligence. When OpenAI and other starts, we didn't know how to do it. And then we found this one thing called a transformer that pays attention to the important things, like you guys are paying attention to the important things that we're saying right now. As opposed to just treating everything the same. And it figured out that we could scale it with these GPUs like Gantyx super computers.

6:47They didn't know exactly how to do it, so they just kept on adding more and more super computers. And we've got a lot like our super computer is 10 times faster than NASA's, right? new super computer is 10 times faster than that. Again, exponential. But then they thought this solves everything, it solves everything. But have you ever seen a generalized system that outperforms a specialized system? No. But that means you have to get people the tools to make it a specialized system. You're never going to send all your internal company data to chat GPT or GPT4. So that's an interesting point. And I don't know if I heard it from you or elsewhere where some of the banks and large corporations are not sending their data and allowing employees to use JATGBT.

7:30They don't want Microsoft having access to it. So what's the alternative? How do you go around that? We have the most popular language models in the world, GTJ, Neo and X, 25 million downloads or know about to release our next generation open source models. Again, base models and then you'll have the banker board versions, Indian versions. So I get to use it on my own system that no one else has access to. You get to go as it via Amazon, Google, Intel, anyone, on -prem in your cloud, you own it. Again, it's like you basically, it's like actually one thing. It's like, again, these are like very talented analysts.

8:03Getting an analyst on succumbent from Microsoft, he goes back to Microsoft versus hiring your own analyst. These are the two modalities. You'll have these big private proprietary models, where results of SQL source and you have these open, interpretable models, where you see how the cookie is made, is it worth? That you own. And I think the latter is far more powerful than the former, because you don't need a model that does everything then. You give people the tools they need and the ownership to create their own experiences and accelerate themselves. Let me ask you, you know, we're all here trying to understand where this is going.

8:35How far out are you able to see? Can you have a sense of predicting where we are in 10 years, five years? I can't see past five years. You can't see past five years. No, because by next year end of the year I believe you'll have chat GPT on your mobile phone. See that one more time by the end of next year You'll have chat GPT on your mobile phone with our internet. Wow. So maybe the talking about someone that yeah And is a chat a stability chat, right? Yes, stable chat Yeah, you know, I guess let's name it properly. Yeah, so I Incredible because you've talked about taking how many terabytes of data down to a couple of gigabytes so like Actually, this is one thing that I wanted to talk about quickly.

9:18The easiest way for us to communicate has been speech. What we're doing now, text has been harder, but you've seen what chat should be. It's not easy. You just ask it to rewrite your kind of emails. And with Office 365 and Workspace. I'll do that. And visual was the hardest. These have all been flattened by this technology now so we can communicate anything. But the technology behind it, as you said, was a bit crazy. To do this, what we did, so stable diffusion was our text image model. We were like, do we do it big and do loads of layers? Or do we make it accessible? And we were like accessible.

9:46So we worked super hard Manhattan project and we took 100 ,000 gigabytes of images, two billion and created a two gigabyte file that can generate anything. Okay, let's slow that down because the meaning of that is insane. How many terabytes was it? It was not a point, 100 terabytes. 100 ,000 gigabytes. And the output was a two gigabyte file That can create anything and that 2GB is you know sits on your phone your laptop In your brain it was the whole stack so intelligence is compression You're gonna go away from this event and you're gonna compress the in things that you learn and kept tea takeaways Right, that's what we do and that's what these machines do now But again, it's a single file four of the top 10 app store apps in December were based on stable diffusion and it was the entire stack a a single file of weights of numbers.

10:40That's extraordinary. Words go in, images come out. And it's the same, chat GPT is a single file, I should say a couple files, but this is a single file. Words go in, code comes out, amazing essays come out. So this is something a bit different, and again, it's like, again, you've taught an analyst just about everything. It's an artist, the analyst, and things like that. I think we'll get it down to 100 megabytes. And it went insane because we made it like that. So this is what you texted me and I said, huh? So a lot of us were doing Web 3 and we were all like developers. Developers are cool, right?

11:19In three months we ever took Bitcoin and Ethereum and develop a popularity. So on the bottom here is the years and on the access and the y -axis here is the number of Get hub developers, get our stars, yeah, developers who love it. So when you said me that, I thought that was the access line. Yeah. So cumulatively, the whole ecosystem built around this thing, which is a bit like a game engine, has overtaken Linux. Again, Linux 20 years this five months, because we gave it to everyone and it runs on your Macbook or your iPhone with that internet. So you have this amazing thing that can create anything.

11:59I think that's profound, right? So we've seen the speed profound indeed. Yeah, we've seen the speed just explode. Can we expect it to get faster and faster? Yeah, I think you can. So look, words go in, images come out. A lot of you have kind of seen this. You can generate anything. But since the release in August, we've spread out 100 times. Wow. We've gone from six seconds in image to 60 images a second. And the quality has improved and basically the next version is photorealistic. So 60 images per second is like video. Yeah, we pretty much got video. So my question is, is Hollywood scared shitless or are they excited?

12:38The scared, it's a mixture. There's a few people who are scared shitless and we'll talk about some of the stuff coming down the pipeline later. Because you know, we turn on our discussion like, should we scare everyone shitless completely? Or should we kind of do hope? And so I'm going to do hope actually. I think I'm going to focus on this. This is the most disruptive thing ever because, again, humans can scale, so you don't need as many humans. There was an MIT study, I think, I'll send it to you, it just came out. It showed that basically the third to the seventh decimal got like 30 % better, the top 5 % got orders of magnitude better, or multiples better, with this technology.

13:10And so it lifts humanity and the ability to communicate and do things. So this is the hopeful future that we share, and the abundance mindset that I think it's really important for people to take away. There's enough fear and some of it is deserved, but there also needs to be hope. I think we always have to look at the unchanging between versus the inevitable. So an inevitable is 41 % of all code on GitHub right now is AI generated. Wow. To six months. ChatGPT can pass a glue level three programmer exam and it will run pretty much on a MacBook or a file in the year. This year, right now. Yeah.

13:48There are no programmers in five years. No programmers in five years. So those are you with kids who you were having, you know, with Python lessons and so forth. Maybe it's instead helping them to understand how to ask great questions or give great directions or prompts. Yeah, like with this, this is a technological model that we sped up a hundred times because we have amazing developers and we have communities of hundreds of thousands. I went to GPT -4 and I said, help me write some code to change the nature of inference something called int8 to int4, which is only a year old, so it's not in the data set.

14:24And it figured it out. Awesome. I also create asteroids in D3. A lot of the high school thing, yeah. And I probably paced it at work, the game straight out. So again, we have to kind of think about this and we have to think, but what can you do? Well, the answer is you can do anything now. Because a lot of the stuff that blocks you isn't there anymore. Any of you can now be creative. Any of you can now build systems. And so you build the systems that adhere to the unchanging demands of people to make their lives better. And that's value and they'll pay you for that value. We're all creators and we're in fact all we had will I am last night at our patron dinner who's this incredible creative energy across all modalities.

15:07And all of a sudden you're not restricted by what you learned in school. you can bring the best to anything you desire. It's best to anything done. You can control it. So you can just say, like, you know, take something and change it to its original form, you know, like this, you know, from mindset to materialization. Which version of you do you like best? I like that. Exactly. And this is just with words. So there's no more prompts. I just say I want to change it to this or that. And it happens instantly now with a new version. You know, you can say, let's do dynamic adaptation again. It happens in one second.

15:42You generate all of these variants, you know, and you don't need those prompts. People say like there's all these magical spells. You don't say that. You say I want to turn Woody into a still from a Western, right, or a place of fruit with cake. So it also becomes more natural. What are interfaces in the future? Interfaces are nothing. It's all about human extension, and it's about information theories at the core of all computer science. Information is valuable in as much as it changes the state. Now, we finally have a friend to be with us all the time that can bring us the most valuable information and the most valuable state changes.

16:14You're a 20 -year -old on our Zoom audience here or in the room or 23 -year -old and you're trying to decide what to do now. What is your advice? You just throw yourself 100 % into this. It's the biggest change in society ever. It'll be more disruptive than the pandemic in the next year or so. and I led the United Nations A &H to be against COVID -19, kayak, you know? Helping organize the world's COVID knowledge and making it accessible and useful. Yeah. So I saw that coming. It'll be for positive and negative, because you can do things that could never be heard of. As an example, we have an upscaler.

16:49It goes from one to eight pixels to 4K in 1 .5 seconds. That's it. That's amazing. What can you do with that? This is the thing. You are given now the recipes and tools. What can you do? So there's going to be a thousand companies spending ten million dollars each this year I spoke to one of the big four accounting firms. They were like do we need auditors anymore? I'm like probably not as many Right and then they said okay, we're gonna spend three hundred million dollars in the next two quarters trying to figure this out but that's times Many many firms everything everywhere all at once and again.

17:25This is the nomencleness We've seen this type of thing happen like with a Silicon Valley bank collapse I've society's been more and more connected and we've talked about expert systems. How long have we talked about expert systems for? A bit yolks. They're here now, right? And again, this is why you can do these crazy things. Like, why don't you create your own studio? So, how do we play this? We play this video. You speak and magically happens. Play. There you go. This is a video of rock paper scissors that was done using our technology by two guys in like four days. And it's a whole seven minute video at this level of quality.

18:01That's all you need. So why not create amazing kind of videos and things like that? But we're going to get to a point soon enough and I'll ask for your prediction where I can say I want a movie about this theme, 90 minutes long, starring my favorite stars. How, when do we see that? That's a nice movie. I think you get there in the next couple of years. Like Rock Paper Scissors too. Again, look it up on Google. We'll be five times as fast. The next version will be fully generated by putting a script in. I mean, like this one, if you press play, and it automatically generates a music lyric video.

18:34So in fact, we don't have sound, but you can go on deep Floyd AI on Twitter and see it. Automatically generated kind of this music lyric video, you know, just from a song input with the entire kind of style of this as well. So everyone here is a filmmaker. I mean, for me, we have a common passion in disrupting and reinventing education, where I can have in a virtual world learn about anything I want a time machine to go and have my favorite characters teaching me about what anything I need. And that's two years away. Yeah, I mean, that I should might be here right now. I mean, again, kind of you seeing Glossis 3D displays and also it's so crazy things.

19:15But one of the things with this business is contextualization. Because again, we are the stories we make. So one of the things we're working on that's quite fun is can you make remake Game of Thrones as a J -Drama from Korea? So you and put in the characters to Korean right well, I'm also trying to get George R .R. Martin and H .F .Y .O. to let me remake Game of Thrones season 8 because it was terrible Yeah, and actually it's very interesting why because Game of Thrones season 1 to 7 Everyone had agency in the meaning was in the interaction between our stories of the characters where as the last one tried to get us to an ending and so it felt so disjointed.

19:50So I'm like, I'll help you write your book, we have the script writing AI coming out and I've looked writing AI coming out. And so we can tell better stories that are more empathetic and engaging. Because again, this is more engaging for the thing. You can customize your thing everywhere. You can customize to every learner. How crazy is that? And there's even crazier. And there's a lot of all at once, every new theme. Everyone, let's imagine you're in the audience here, you're a business owner, entrepreneur CEO,

20:17and what do you do next? I think what you have to do next again, it's kind of what you said. You need to have a dedicated, there's nothing more important in your entire business than this. I have to, so I wanna echo that, right? Today, the world changes. Electricity is made available, the telephone is made available, the internet's made available, and this is all of those things together, all at once everywhere. Yeah, this is 5G, this is bigger than 5G. over a trillion dollars went five or two more going to here. Actually, I'll give you an example because again, it's always the unchanging versus the inevitable, right?

20:50I gave the whole of stability AI the weak off at the end of the egg, because I was like, how do you craft your doing a lot, right? Because we do everything, right? And we do, we only have 150 people. Yeah, we do every modality, because we have the power of community and open behind us, saying the standard. So I give the whole team the weak off said, go to sleep, turn off all the things, because you're going to die if you don't get some rest. So of course business side of covered that stability I see your mom must act competitive to open the eyes is you're all going to die in 2023 We might me make press as well.

21:23What can I say? But then what happened is I you know just chilling and I got like five calls from head masters of UK schools Emma and what's our generative a eyes strategy? What all our children are using Having a chat GPT to do their homework, essays, I don't set homework essays, get good. So again, homework is never the same for every teacher all at once. And so you have to look at your industry and again, what will not be the same when you have the ability to scale humans? This is the question that you have to ask yourself. Like an eaten now, a bastion of British education, they do their essays live by hand because I told them to do that because I went to Westminster who I told to embrace the technology.

22:04I'm just a bit of a writer, right? So again, it says Peter says, the only company is the Embrace -Anne, those who don't. I think the way that it pays out, because I was a hedge fund manager invested tens of billions of dollars, like Einstein Mark as well, is that if you are in something that's a SaaS company, you're gonna have to really think hard, because GBT4 for example, and again, we'll have our equivalent of that, but we're taking a different approach of swarm intelligence versus general, has a 32 ,000 token context window, which means you can feed it 20 ,000 words of instructions. Well, what is OpenAI?

22:41That's their one, GPT -4. The one that they use. OpenAI is the best language. All you should use OpenAI, and you should use our models coming up as well. Private data and general data. Hybrid AI. But that means you can basically tell it all of HR instructions for something like Workday. What does Workday do then? Because it can be replaced by a single 200 megabyte file. Which you can query on your laptop or your phone? Well, yeah, eventually, but not now. But it doesn't matter. It costs like $2 for query, right? I mean, it's insane. But then you have to think, if you're an arregulated industry, how super normal margins?

23:15You know, if you are in a creative industry again, you can basically either embrace it to have higher revenue, cheaper costs, or not. Like I give you an example. Well, one of our maybe directors did a film, well, actually I can't give you an example. Let's just say millions of dollars of being saved at the moment, using this technology. And again, you just have to really think through it. So going back to our audience here, who use technology but don't have AI embedded in their DNA, I've shared the idea of having a chief AI officer who's really a strategist for you because you're not building your own AIs, your deciding which platforms to use, right?

23:53Yeah, and again, a very few comments will build their own AI. So I think it'll just be Google, Microsoft, Open, I and us that build the foundation models that everyone uses. Because you don't need to. Because again, it's like where the university is churning out the most amazing graduates effectively. You know, some of them you can borrow and they're specialized by those guys and some of them you can hire yourself through your GPUs or kind of whatever. My suggestion is that that... What's the key, but the thing is, how do you find that person? You know, until they're in AI. No, how do you find that person?