Best of 2023: This Is One of the Biggest Economic Impacts of All Time

27:45And then we distribute that through our hyperscaler partners like Amazon and others. Like Snowflake made$900 million via the Amazon sales channel last year. and our value is in standardizing and making that predictable and stable, shall we say, across all these modalities. So for the base layer we use, we get our revenue shares, we get our customized things on the back of that as well and we can be open AI for everyone else because there's no other company that can guarantee that you can build a model. So we have all the top companies in the world coming to us and saying, can you create custom models for us?

28:19And we're like, yeah, how much, how expensive is that? It's very expensive. compared to seven weeks ago. So it's a mixture of the standardization and this customization. That's where we fit in for all the private data in the world. If it's public data or if it's generalized stuff, you can use the APIs and you can send your data to Google and Microsoft. There are also going to be private instances of ChatGPT and others where they guarantee not to look at it. And that's, again, a great halfway house. And so the landscape is these private instances, these generalized APIs, and then the private side, the standardization layer is us.

28:56So you talked about the specialist models. Where are you now with all of the suite of products that you've got? Because you've got a lot and more coming. Yeah. So we've released the image models and we have the next generation of that coming out. They're cutting edge for controllability now. So we figured out fingers. We figured out text. We have video models training on 11 ,000 videos. When is this video shit coming? Because that's, I think another big shock that people don't realize what is about to happen. I think the first release is in a month or two from us. We've got an amazing team. And this is text to video.

29:31Text to video, yeah. You type it in and generate videos. You can already kind of just act a video and do style transfer on it, like the Corridor Digital videos, and that'll be heading towards real time where you can turn what we're doing right now into full animation, almost real time. But then you want to be able to create anything just by describing it. which is coming. We've got script writing technology as well. So I'll write scripts in the style of some famous directors we're about to announce to help the directors. The whole of media is going to change as a result of this. Yeah, we're spinning at Real Vision to try and just figure out where to play your cards because unless you're a big company, you've got budgets and you need to figure out what the hell to do.

30:15Well, I mean, that's the thing. You can auto-generate a podcast based on the intents of the person on the other side, right? With a human realistic voice. And then even Raul. You auto-generate entire interviews back and forth between people. Yeah, because there's enough content of me online that you can train a model. Yes. And so you can have a customized Raul every morning for every single one of your listeners. I think within a year, definitely two. oh my god how do i keep up i mean it's just really hard um because you don't want to get you don't want to get disrupted because somebody's going to do it that's the the terror in all of this is there's some 24 year old going well fuck it i'm starting from scratch and i'm gonna yeah like there is an infinite seinfeld channel now um some of the jokes are quite funny you know there's this thing where they've got the dynamic defects again there are certain things you have to always think where is my edge in this new world because what happens this is a regime change right it's a regime change in the way that information flows around the world and it's going to restructure the value landscape and then people find their own areas like where is our value our value is being the standard for open models but what do we mean by open we mean transparent free-range organic you know how they've been fed you know what's inside it and there's a whole ecosystem and support around it right so transforming the world's private data is our thing.

31:43And that's where I'm occupying. Because I think that's a hundred billion trillion dollar company. That's something I want to IPO ASAP. Other people are finding their edge elsewhere. So like I said, in regulated industries, if you adopt this technology and you can scale the human component, boom, your costs go down dramatically. In media, if you use this, then you can leverage your existing IP and boom, away it kind of goes. You can surface stuff. You've got intelligent interns and analysts and associates and VPs is coming. You don't want to be computed away. At the same time, stuff is slow. People still use Lotus Notes, right?

32:191.5 million people are still on AOL. Everyone said that meta was going to disrupt Google search. It never really happened. Hopefully, they'll turn back to being Facebook soon. Yeah. Maybe they'll change their ticket to AI. you know so okay so we've got text to video we've got script writing we've got the next version of your of um language models the step of fusion so we've got that's the the kind of gorgeous ability text to image then you've got language models talk us through the your version of chat gpt and how you play in that sphere because that's interesting yeah so you know existing language models from our Eleuther AI community, generally 25 million times.

33:04So they're basically standard in language models already. We don't really communicate that much. In fact, we've just spun out Eleuther into an independent charity, or 501c3, to do interpretability, evaluation, and alignment, because we think it's important to have an independent entity rather than us going and doing it by ourselves. And we're bringing more and more parties to that. So the language model is the first bit. That's the bit that's trained on the corpus of the internet. And we have our new data set, the PAL version 2 being released in like a week, right? And how recent is that versus chat GPT, which is 2021?

33:35It's like a few months ago. But again, these models now we're teaching to be auto-updating and things like that. Because the other part is we have a code model coming. So we released the first version of that last month, 6 billion parameters. And so when you combine a language model with a code model, you can do dynamic SQL database lookups to fetch recent data. So u.com, Niva, and Bing do that. to make sure they're up to date. So you have this generalized knowledge and then a lookup of recent facts to compare against it. You don't have to have one model to do everything, although eventually you might be able to.

34:08So the base thing is first the language model. So our language models are very good. That's good. Like melting lots of GPUs. But then the next stage of that after that is you instruct it to human preferences. So that's reinforcement learning with human feedback. Because you have this generalized thing. It's like putting someone in front of a TV and taping their eyes open and they just watch it. They learn about everything, you know? But sometimes the answers are a bit crazy. So you teach it what type of questions and answers via a reward function humans usually like. That's the RLHF bit. That takes a few months, right?

34:42And then you take sector and domain-specific questions and instructions and you add that. And then you see how people use it in real life and you iterate and you adjust the outputs that way. So, you know, that'll take like, I don't know, six months or something like that, the whole process. but the language models are the first bit, and then you move forward to that. So it's going to be like a stability chat. Is that how you think of it? Yeah, we've got stable LN first, which is the language models, and then we've got stable chat, and then we've got stable code. So people will use stable code for their internal code repositories, stable chat for their internal things, and then they'll pull in Google and Microsoft slash OpenAI APIs to create amazing hybrid AI experiences, right?

35:22So you can use both. So you can have your private data, private model on stability, but you can still use the generalized Google one or whatever you want to do. Yeah. Like you could have Google Slides, like I said, and it pulls in your private knowledge, and then it pulls in this really amazing script writer from OpenAI. And boom, you have auto-generated slides that are beautifully cogent. Or you have it in Office, et cetera. I think that's the future. You have private models, and you've got public models, right? Just like you had on-prem and cloud. um so that's why we're building models for every single type of media because every single type of media will need their own models and then you bring them together what do you get whole worlds created dynamically movies created dynamically you know and isn't isn't this just the the metaverse moment as well because everything just goes into that now because you digitize everything you can create.

36:21Raoul can exist in five different places at the same time, talking to different people in my own voice and all of that kind of stuff. RAOUL PAL Yeah, this is the breakthrough technology that Facebook should have focused on or meta. They did not. But now they're lucky because they get this. And so what we're doing by building open models that everyone can optimize for and standardize around will drive the acceleration of bringing that multimodality, those digital twins and other things for private data. Because otherwise, what happened is you'd have this Microsoft choke point on the internet, just like it had before, right?

36:56Now that at least there's another option and people can own their own models and own their own data and have localized versions of it, customized versions of it, rather than being beholden to that. So that's where I realized the gap was and that's where i've gone with stability um i think that's a better place to be as a business i hope well i know um and also it's good for society because again you don't want this technology to be controlled by just a few we want it to be widely available so let's get philosophical there are two outcomes one is it accrues to some two or three superpower companies albeit it's given to everybody and a combination of the two is what's happening so it becomes kind of ubiquitous there's no stopping it right your choice of stability AI as an kind of open source means that it is now essentially viral yeah I mean like you just look at stable diffusion most popular open source software ever In three months, it overtook Bitcoin and Ethereum cumulatively on GitHub in terms of popularity.

38:06It took them 10 years. It's overtaking Linux as an ecosystem. Putting this out there, people will take it, extend it, and build around it. We've seen 16-year-olds to 60-year-olds kind of doing that. So our image models are by far the best. Our language models, once they get going, we believe in a year, will be far the best. And they'll be accessible and usable. Because otherwise, you're going to have this digital divide, right? On steroids. you'll have super augmented humans who have access to chat gpt and those who don't again i think we mentioned last time like i'm usually very supportive of this industry but i spoke out against open ai because they banned ukrainians in ukraine from dali too the image thing i'm like how can you do that like i think i don't say unethical very often because i think ethics are complicated but i was like that's just wrong you know they recently changed that in the last few months, but there was a long time when no one in Ukraine could have superpowers for art.

39:00You know? And I understand there could be reasons around that, but you can't have a world where you have these restrictions. There was just a study released about ChatGPT. Guess what? ChatGPT's ideology is that of Silicon Valley.

39:16Okay? You might agree that that should be the only ideology. my thing is that the world contains multitudes. And so we need to have every country will need their own national models to reflect their culture. And we need to distribute this widely as infrastructure. So again, this is why we have Imagine Worldwide. Bring this technology to the kids first and let them be the best of themselves, as it were. And when, in your mind, did these generalized models become smarter than humans? I mean, Mo Gorda, I don't know if you read his book, Scary Smart. So he was the guy who ran Google X, a really interesting guy.

39:52He's just like, end of this decade. Well, I mean, look, half of people have below average IQ, right? Around that. By definition. But 95 % of people think they have above average IQ, right? The barrier is not high for generality, unfortunately. Well, fortunately, I'm not sure. So I have no idea. I just know that ChatGPT is a better coder than I am in certain instances, right? Stable Diffusion is a better painter than I am. Our music models are better musicians than I am. I'm sure that MedPalm is a better doctor than I am. And do these all become just one big massive model in the end? I think there's millions of models.

40:38Yeah, I think you need millions of models because as you have a model that's specialized, it can be highly performant, right? But these models are coming to the edge. If you put all of the specialized models together, you get generalized AI. You get generalized AI, yeah. So my take on the AGI debate and all that thing is that swarm intelligence, augmenting humans, is the only real alternative to an AGI that overtakes everything. And I think this can be used to create swarm intelligence. On the way to that, you make people happier and you give people agency, which I think is a great way to do it.

41:09because I don't like this idea of just building a multi-trillion parameter model on billions of dollars, where people talk about alignment to these models with the immersion properties, right? So they're like, we can constrain it. What about the alignment discussion? Because that's important as well. Yeah, so alignment is that as we build a model that's generally more capable than any human and can learn to be even more capable, we can align it so it's beneficial for humanity and doesn't wipe us all out. because it's like, man, what's the easiest way to make humans not sad? Get rid of them all, right?

41:42But for me, that's what's prognal to freedom. So if you have someone more, we've all had people more capable than us, right? Can you constrain them? Yeah, if you restrict their freedom, and that's the only way you can do it because they're by definition more capable than you. And that terrifies me. So for me, one of the reasons I want to get the technology out there, I want to get it where it's useful in education, healthcare, creativity is so that we can have dynamic form intelligence. And then that can coalesce to maybe counterbalance that, which is a crazy thing to say. But again, we're seeing it live.

42:17Having these conversations a lot now is like, it is not certain what that path is. It isn't. And so the common refrain that I get is two things. The only way to beat an AGI is to build another AGI that stops the bad AGI from happening. That strikes me as being very bad, right? The other thing is, if we don't build AGI, the Chinese will build AGI. And again, that seems a bit silly. So my thing is, again, get this technology out, create standards around it. One of the things we introduced was opt-out of our model data training sets. Nobody else has done that. Now, is it because they think it's the legal requirement?

42:55No. Moral requirement? Not really. Because again, if it's public and open, and you're using the scrape, I should be able to for that. But I think it's the right thing to do. And so setting standards around that, pushing back against governments who actually want to be even more permissive about what we can do. I think we need to set standards like that. How do you stop somebody, Israel, or somebody saying, well, we're not going to adopt any of your standards, therefore we can capture market share and we can do that? I don't know. I think all you have to do is just have to try and take a lead and make it the defaults.

43:23So again, the models that we are training, how many people have really stable diffusion models, even though they have the training code. And it's like only a million bucks to train. Again, compared to the value created, like Lenzo is making two and a half million dollars a day on the app store. That's just one app, right? From this technology. But now would he train them? So being in this open, transparent standard, like we have quite an important place in setting the standards of what is reasonable. And then if people don't adopt that, I said, maybe they outcompete. I don't think they do. There are elements around that, but it's hideously complicated, unfortunately.

44:01And especially at the pace, there is no time to breathe, unfortunately. And it's really interesting listening to Sam Altman speak and seeing his tweets. I mean, he's fucking terrified. They said that the AGI alignment plan, and it was like, we're going to treat this as existential, even though some people disagree. I think that's incorrect, because if it's existential, you wouldn't build it, unless you're terrified of someone else building it. So there's a very unpleasant race dynamic. That is the case, right? there is game theory here yeah it's kind of a known unknown i think you know it's an incredibly difficult thing we don't know if it's 10 years 20 years one year it's a known unknown as it were because again it's when do you have generalized intelligence my take is stop trying to build generalized intelligence you don't need it because you can have superhuman narrow intelligence already that makes us better and we should focus on that and again getting out there the hive mind how people use this and then take it and extend it, especially if you standardize the base.

44:57People optimize around that. That's what you've seen with stable diffusion. That's what you're going to see with the language models. That can outcompete the deep minds and open AI's and others of the world, I think, in maybe getting there first in a small manner that is safer. But I don't know. And where are regulators in this? I mean, they can't even regulate crypto. How the hell are they going to regulate this? I mean, look, on the Section 230 discussions at the Supreme Court, they're basically getting around to the internet right now, right? I mean, literally that's in the comments. You have to help them, right?

45:31And so I think, again, regulation should be introduced, but you need to help the regulators because you can't introduce it in time. And again, the regulation is actually pro-competition now because every single country is looking and looking at other ones like, holy crap, who's going to adopt this technology faster? It's a very unfortunate race condition. so this is why i said like this is one of the biggest economic impacts of all time i just think it's going to be bigger than the financial crisis bigger than the pandemic there's a lot of irreversibles and some key questions will have to be made around this this is also you know like from my side i'm like i need to make stability a public company because we need to be accountable given our place in this also because there's going to be a supply demand imbalance like no other if i excuse it because i was about to ask the question there's basically no way to invest in this right now for the general public.

46:17Yeah, you can buy semiconductors and you can buy Microsoft and Google, but that's it. That's pretty much it. And even with Google, you've got the super normal margins that are definitely going to come down as a result of this, right? And they'll get punished for anything. You've got Photoshop and Adobe and things like that that will get a benefit. But then what happens to the competitor comes in? You've got Nvidia, it's gone high, but what happens when people shift to ASICs that are 10 times cheaper? you know so there's no easy place i'm like i'm gonna make an easy play um and then that will hopefully allow me to you know distribute the governance and have a positive impact on this even while making it so that we can remake game of thrones season eight because it was crap and things like that dynamically so again there's this terrifying aspect of this but there's this amazing aspect of this yeah educate every child what if this ai kills us you know what if someone gets it before.

47:10So it's this duality that I deal with every day. RAOUL PAL Yeah, it's like nuclear weapons and nuclear energy. RAOUL PAL Yeah, it's like, what if you put the nukes on the bottom of the rocket, you would have got to Mars already, right? Or you put them on the top. RAOUL PAL Yeah, it's a very good point. The other thing that is coming down the track, and I know Google have been working hard on but also concerned by it is quantum. Because quantum then increases the speed of this thing like to Reed's law. It just goes exponential, exponential. Well, I mean, so you've had the move. What happened is that you couldn't, with the transformer architecture that paid attention to the important parts of things when learning principles, you could scale it with these A100s and things like that.

48:00But after 1 ,000 or so, it tailed off just because of thermodynamics, actually. Like I said, we're literally melting GPUs now, and we've got like 6 ,000 training our language models and things like that. It's like a half a billion dollar computer. That's actually been solved now, so that's linear scaling. You can actually emulate quantum qubits on 8100s, and you can scale it pretty much out there, which has big implications if you think about it, because people are going to drop billions on these things now, where food said only drop hundreds of millions. As quantum computing comes in, the key thing here is you have these models forming one part, and then quantum computing is great for optimization equations.

48:39Because one of the things you've seen is you've moved to deep learning as a key thing. Now you're introducing reinforcement learning. And this brings up forward things like AlphaGo and those AIs that beat humans at StarCraft and stuff like that. And then you bring quantum into that, you can basically optimize complex systems, which is kind of crazy. yeah how the hell are you staying on top of all of this I know you're at the epicenter of it so I get that it's difficult man like every day there's not a breakthrough seriously when you look at ML papers on Arquib it's an actual exponential it had a 26 month doubling it's now like 13 or something like that and again it's parallelizing RAOUL PAL A lot of people ask me, where's the financial markets people with this technology?

49:33Whether it's Renaissance, I don't know what models they had, because they were obviously doing a lot of stuff early on. Where's that coming? How unfair advantage do the people who can build the models get for a while? RAOUL PAL Well, I think they get a big advantage because you can understand how to break down the markets. I think you move to AI-based market makers sooner rather than later. There's a whole bunch of stuff. the general sophistication is very low because there aren't many people that understand these technologies. There's only a couple of hedge funds I've seen and I'm not even doing it properly.

50:05Like if I have this time and space, I'd love to learn a hedge fund. I don't know. So it's going to come. And again, I think there is an unfair advantage from being able to do this. Even more so if you have more degrees of freedom, shall we say. Like what if you could read the stories and create the stories? We're going to see agents like that that understand how things spread based on the variety metrics. So I think markets have already moved highly narrative-based, if you look at it, like more than kind of other, like what are fundamentals anymore even compared to narratives? I think that will just continue to be exacerbated by this technology.

50:44How is, and one of the things you might have seen me talk about this, and we briefed on it before, is how the hell is society going to deal with the deep fakes and the scale and speed with which we can create that going into the US election in two years? Well, next year, right? Next year, yeah. Basically got perfect deepfakes that were coming on away. And it's pretty much real time. And the voice is like the audio DJ that just got released in Spotify. That was my sister and most technology. Perfect voices, right? Dynamic with full emotional range. I think the only way is you have something like content authenticity.

51:21you have this kind of a mutable trace from a curated source because then social network becomes even more important with the way that you curate these things because detection is very important. Like we're going to do a$200 ,000 deepfake detection price for open source, and then we would like to be a waste condition. So maybe it'll be some sort of blockchain. Maybe it'll be something like content authenticity that does an encode metaphor. So there needs to be standards sooner or later. But I think already people are like, I don't really believe that anymore because I just saw it on TikTok. up.

51:49I'm pretty sure Tom Cruise isn't doing all that stuff. RAOUL PAL Yeah, it's fine in normal life. But when it comes to points like pandemics, things of societal importance, that becomes complicated. I've actually spoken, spent some time speaking to Google, Facebook, Amazon, LinkedIn, all of this about authentication of people as well, because maybe you just authenticate people. This is why Meta is introducing this people authentication. This is why I said, if you look at this variety and spread, it's all about the curation aspect. And so there's a verification thing, and that could be blockchain or it could be content authenticity.org.

52:28But then there is, what are these key choke points on information flow? And what do they look like? So you were working, and we've chatted about it privately, on cool stuff that are going to make people go, oh, my God, again. When is some of this cool stuff that you're seeing coming out? Or is there any cool stuff that's out already that you think people haven't noticed how amazing this is? Well, we've just hit 4 controllability on image. So you can take the pose you are now, and then you can basically transform yourself into anything, move it to 3D. There are new types of interfaces. How long does something like that take?

53:08Talk me through that process. So I put in an image of me or a short video of me. I put it into this model. And then what? And then, you know, you can transform yourself into a transformer in like, I don't know, a minute or something like that. And then once it's trained, it takes a second each time to transform yourself to anything else. But then you can export your 3D model, adjust it with a Blender plug-in, and then you can just do that. Or you can take a whole movie and then transpose it with coherence probably in the next few months into like an anime. You can make yourself Dragon Ball, whatever.

53:39then you can view that on I don't think I showed you one of the companies I'm an advisory board of they just released their glasses free retina quality 3D tablet so it's 3D without glasses and it's retina quality I'll send you a tab and basically we convert 2D to 3D live as well and you're like what is this magic craziness especially in advanced technology is magic there's too much magic out there the question is you know is it destructive or creative magic it's a bit of both unfortunately it is yeah we just have to be honest with ourselves we just don't know but it's not going away you can't put the genie back in the bottle we were always going to somebody was always going to build this yeah someone was always going to build this and that's kind of what i saw and that's why you know stability is like 16 months old i accelerated this so so hard and now i'm accelerating it again because like again we've got to have options right it can't just be like controlled by a few choke points that okay they get economic excess but at the same time like are they doing the right thing i don't know but there should be options am i doing the right thing i don't know but at least other people can take the technology and kind of use it hopefully for their own customization and the ownership structure is different is there a good business on that i think there's an amazing ridiculous business from that as well and the whole entertainment industry as well, completely changes.

55:07Oh, here's a question that somebody asked me the other day, and it's a very valid question, is has the porn industry adopted it yet? I don't know. You know what I mean? They've always been very fast with the adoption of technology, because then that changes a lot of the equation. You can see these AI models now, so modeling, so you can get whatever kind of model you want to wear, the clothes that you want. It's like, oh okay well i mean like i'd say my hope is probably that the exploitation right in that industry decreases if you have digital variants of it um but i'm not sure how it's going to be adopted at the moment i do know that just like you know real life still life pictures some of the kind of models that included not safer work actually had better anatomy and things like that because it learned again it learned from still lifes um and nudity and things which is kind of crazy to think.

55:59Again, I think any information, media flow based business gets changed by this. And that is basically Western society, right? Like, I can't think of very many places where this isn't applicable to save money, make money, create new experiences, etc. I can't get my head around it, because it's literally every single thing we see, do, learn. I mean, just it's everything. It's everything. And there's this question, like, I'm reasonably convinced that in two maximum five years, you'll have chat GPT on your smartphone without internet. What does that mean? When you have a very capable analyst level person on your phone, Bing is no longer stupid, Siri is no longer stupid, and it learns about you.

56:51But it also means that the Internet of Things can have the same. Yes. And they can be embedded in your fridge. Yep. And it can go everywhere. I said you'll have it in your toaster, double diffusion in your toaster. Why? Because then you have fancy toast. Fancy toast. Of course you will. I mean, you know, the Internet of Things is headed that way, but once you give it AI, it can do all sorts of stuff. Well, Web3 was missing AI, right? I always was very puzzled by that. You have identity, but you don't have AI. So again, I think people will adopt this faster than, well, they are adopting it faster than any technology we have ever seen.

57:26And we've been in markets for decades, right? Never ever seen anything like this. Nobody has. Because there's an installed base and it just seamlessly fits in because it takes structured and unstructured data back and forth. And then he said, like, eventually it will be on the edge, you know, and it will be available to everyone. And I think that is an incredibly uplifting, agentic thing. or it's the panopticon and all panopticons. I don't know. Yeah, I don't know if my fridge is going to take over the world or not. But once it's off the internet, you can't control it. Yeah, there's a great story by I think Corey Dotswar called Unauthorized Toast.

58:02The intelligent toast maker, you don't pay it, it stops making you toast. You know, you've got Hitchhiker's Guide to the Galaxy with, what was it, Red Dwarf with angry toasters and things like that. Oh my God, it's going to be a crazy future. I don't want toasters getting the attitude. RAOUL PAL, I'm getting up with that already. RAOUL PAL, Yeah. Emad, listen, amazing again to catch up with you. Thank you for giving us your time and good luck with everything. Let's see where we get to. This is crazy. I mean, we only spoke about a month and a half ago. EMAD DURNER I don't know. RAOUL PAL, Titan compression technology.

58:33RAOUL PAL, Like, oh my God. RAOUL PAL, Where are we going to be in a year? I don't know. RAOUL PAL, I don't know. RAOUL PAL, I have no idea. I have no idea. Anyway, my friend, great to see you as ever and good luck with everything. RAOUL PAL, Cheers, buddy. We live in interesting times. RAOUL PAL, We do indeed. There's almost too much to process in this video. I can't get my head around it. It's moving too fast. There is too much happening in too many spaces in every single aspect of everything we do as humans on earth, from creating art, to writing scripts for films, to the medical practices, accounting, lawyers, media, literally everything.

59:17Any single absorption of knowledge has changed, but at a rate of which none of us can understand, A, how to even catch up, but how to even understand it. And when I ask Emad, who's at the bloody center of it all, where it's all going in five years time. It's like, fuck knows. It's moving so fast in so many unique ways by so many people that we really have no idea. But I think Ahmad confirmed my view that this is literally the biggest thing that has happened on an economic and societal level in history in the shortest period of time. And the ramifications of this will live with us forever. Read all the manuals and text functions in an fucking ocursey, Thank you.