Pod Insights Real Vision: Finance & Investing The Exponential Age - "This Is the Most Important Interview in the History of Real Vision" About this episode Emad Mostaque is democratizing artificial intelligence. The hedge fund manager turned entrepreneur is the founder of Stability AI, the company behind Stable Diffusion, an image-generating algorithm that creates new, never-before-seen pictures. It’s the most advanced technology in AI, and Emad is bringing it to the masses. In the first episode of our ‘Best of 2022’ series, Emad shows Raoul how quickly AI is progressing, what that progress means for our future, and why he chose to make Stable Diffusion open source. This type of technology should be available to everyone in the world, Emad believes, not just corporate tech giants. His vision is AI of, by, and for the people – and with that comes great opportunity. Recorded on Wednesday October 12, 2022. Learn more about your ad choices. Visit podcastchoices.com/adchoices
Transcript 11,863 words · 3 speakers
0:00 Speaker 00 The title here says it all. Raul was so blown away by this discussion, he declared it the most important interview that Real Vision has ever seen. In the first episode of our holiday Best of Macro series, Raul speaks with Ahmad Mostake. He's the founder of a company called Stability AI, which is focused on democratizing artificial intelligence and bringing the most advanced technology to the masses. Enjoy.
0:25 Speaker 01 As you know, I have a thesis about the exponential age that we're about to go through the fastest pace of technology the world has ever seen. That's what this whole series is about. This next interview with an old friend of mine is probably one of the most important single interviews you will watch. I can't stress enough how important the concept of what is happening and the speed of which artificial intelligence is going to change humanity. It is something that we all need to pay attention to, understand and use in our favor because it's happening and it's happening at a scale of which you cannot comprehend. When change comes, opportunity abounds.
1:10 Speaker 01 We're about to enter a period of the fastest pace of technological change in all human history, something we refer to as the exponential age.
1:22 Speaker 01 And Real Vision is going to be your guide to this incredible future.
1:32 Speaker 01 Ahmad, fantastic to see you back on Real Vision.
1:34 Speaker 02 Good to be back. Thanks for having me.
1:38 Speaker 01 You were known in the past in various guises on Real Vision because this is not your first rodeo. You were the frontier guy and we're going to talk about a new frontier today. And then also in the pandemic, you helped guide us through some of that too. Give people a bit about your background because it's a fascinating background and it's an interesting journey to where you are today.
2:01 Speaker 02 Yeah, it has been a few different skins. No, I was a hedge fund manager at the start of my career running global macro. Then switched to emerging markets, frontier markets and tech. At one point, I was one of the larger video game investors out there, which is also quite fun. Took a bit of a break when my son was diagnosed with autism and I used artificial intelligence to repurpose drugs to make him better. While advising some top hedge fund managers and governments on various things, including counter extremism and fun things like that. Went back to being a hedge fund manager, did okay. And then, yeah, had a look at the world. And you know, what we do as fund managers is we look for arbitrage opportunities. And we look for where things are building up and breaking down. And it struck me that a lot of stuff was breaking down.
2:45 Speaker 02 And I decided like, let's try and make it better effectively by combining some of these systems and some of these arbitrage opportunities. So one of the big things was education. Another big thing was leading the United Nations AI initiative against COVID-19. I think the last time I was on Real Vision was February the sixth, twenty twenty. And we had a nice chat about how the world was going to change dramatically. And it did. And then I was like, oh crap, I have to do something about it. So we made all the COVID research in the world free, pushed that. And then I was lead architect of Kayak. Which was the UNESCO, World Bank, WHO backed initiative to use artificial intelligence, take all that knowledge.
3:26 Speaker 02 And compress it down to reports for policymakers so that we can be a bit more coordinated. Similar to how we take all that information, fund management, compress it down and try to figure out those key drivers. It was a mixed success, but I think it did some difference. And now from that, something a bit bigger, which is AI for everything, basically.
3:45 Speaker 01 So AI has clearly now become your thing, right? This is your focus.
3:49 Speaker 02 Yeah.
3:50 Speaker 01 So before we get into what you're doing now. What the hell gave a hedge fund manager the thought process that, hey, I can help the WHO and everybody else with artificial intelligence?
4:03 Speaker 02 Well, it's because to be a hedge fund manager, you have to be a bit of an egomaniac, right? You have to believe that you're right and everyone else is kind of wrong. Look, I mean, all of what we do is information classification, and that's the nature of artificial intelligence. So information comes in and then Claude Shannon style information theory. Information is only valuable as much to change the state. So you're always looking for something on that graph or that news thing that will cause you to change from a buy to a sell or an increase or decrease in your positioning, right? I kind of realised over the years, though, like my background was mathematics, computer science originally, and so quite quantitative, quite analytical, that the next wave of AI was something a bit different that would enable us to do something a bit different because we moved from this big data age,
4:47 Speaker 02 you know, where you had massive amounts of information and they extrapolated it and targeted it to kind of sell Raoul some suntan or whatever, you know, to something a bit different, a big model age.
4:57 Speaker 01 You're jealous of my tan.
4:58 Speaker 02 Whereby, I'm jealous of your tan. I'm in London, you're in Cambridge, I'm turning pasty, completely. I have to fly out there, too. So you moved from that area, it was about extrapolation and the individual, to something new in 2017, which was a big model age. AI went from being able to just do extrapolation to being able to pay attention. In fact, there was a seminal paper called Attention Is All You Need about how to build an AI that paid attention to the most important parts of a sentence, the most important parts of an image, to do principle-based analysis, which is insane if you think about it, but it's exactly what we do, right? We have this heuristic.
5:34 Speaker 01 What do you mean by principle-based analysis? Sorry.
5:38 Speaker 02 So rather than doing extrapolation, so rather than doing momentum, beta, AI became able to do alpha. It became able to basically come up with principles that allowed it to understand the hidden layers of meaning. With things. So you can go into a bit more detail, but this is how the human brain works, right? We have two parts of our brain. The future is like the past, and that's a tiger over there in that bush, right? It's the type one, type two kind of thinking. So AI was always one of the extrapolation. And now we have a new type of AI that mimics the human brain by figuring out principles from highly structured data, which is kind of what we did every single day as the finance guys, right? But at a completely different scale.
6:24 Speaker 02 I don't think even now people have appreciated how big a change that's going to be to just about everything. I think we're starting to figure that out now.
6:31 Speaker 01 So before we get into the guts of this, let's go on a bit of the journey of AI. Cause you've talked about where it was, and then we started to see stuff like DeepMind and then GPT three, OpenAI, all of this. So if you can just frame it for people, because a lot of this is going to be new for people. And I think it's incredibly important people understand what's happening, the speed at which it's happening, where the hell are this is going?
6:59 Speaker 02 Yeah. So if you look at DeepMind, um, they are famous for many things, you know, Demson's done a fantastic job. Um, one of the things they did is thinking back, you had Gary Kasparov being beaten by Deep Blue. So you know, us, us older fellows, we remember that, not let the young with a snap and see, never seen a bear market in their lives. Uh, you know, kind of changes happening. What Deep Blue did was it did a analysis of every single game in the past and then extrapolates it and then it brute forced it. So Gary could only think X moves in the future, it could think X plus two, X plus three. And that's how he got beaten by brute force. Now, the Chinese game of Go, Chinese checkers effectively, people thought you'd never be able to brute force it because there's too many things that you can do.
7:44 Speaker 02 There's too many moves on that board, right? Um, and so they're like, ah, that'll never be beaten by a computer because you need to build a computer that's the biggest computer ever, times a million. What DeepMind did is that basically they came up with a self-supervised learning algorithm that learned how to dream, that's probably the best way to kind of put it, about how one would play Go in a principled way. So it didn't even have all the past games in its memory, it just played against itself and tried to figure out how to do this. They made a whole base of these agents with reinforcement learning that you gave it an Atari with no instructions that learned how to play Breakout and then StarCraft and all sorts of other things.
8:27 Speaker 01 Without any instructions, people need to understand that it figured it out.
8:32 Speaker 02 You put a computer in front of something, it just figures it out and then gets better and better. So when you see it playing Breakout, it's suddenly doing these crazy moves like that. And so Lisa Doll was like, I think a seventh dan in Go, so he was a Magnus Carlsen of kind of Go effectively. And it wasn't a current Magnus Carlsen type situation, it's very difficult to cheat in Go because you didn't have computers to help you, right?
8:52 Speaker 01 We're not talking about anal probes in this one, we're talking about...
8:57 Speaker 02 Yeah, conversation's going many ways, right? So what happened is that the next highest person was like a fourth dan or something. So he's like Federer, Carlsen rolled into one, that much better. I said, never beat him. He won one game, the computer won seven. And everyone was like, holy crap, what is that? Because computer learned to play in a completely different way. And it was like, this is playing with an alien.
9:20 Speaker 01 Yeah, because it didn't learn past games. It, as you said, it learnt in an entirely new way.
9:27 Speaker 02 Literally, it played against itself, it dreamt, so then it plays against itself in its memory. And then it does that again and again and again. And so within just a few weeks, it outperformed him. And then these models got better and better and smaller and smaller. We haven't seen the generalization of the models yet, but they're coming. These agents that can basically optimize when you don't even tell it the rules, which is kind of insane. Then, on the other hand, you had what was known as foundation models, which is where this attention-based architecture came in. So in the attention thing, they were seeing where the games are. Actually, yeah, a final thing about this, and you can see this in the AlphaGo documentary, that's on YouTube if you want to see more details, is that as a result of this, did Lisa Doll go and say, I'm hanging up my Go pieces and I'm going
10:11 Speaker 02 to pack them away? No. He started playing against it and now he's even better as a player. And the entire Go... Competitive scene has improved because now they're figuring out brand new ways and kind of gambits and things they'd never seen before, which I think is really interesting. Then you had kind of this other kind of big area, which was kind of some of this deep learning area. We have these big corpuses and you know, like I said, you can do extrapolations and stuff, but you don't understand meaning. And so this is where these attention based architectures came in. It's actually quite interesting. So when I was taking a break from fund management to work with my son, autism is a very interesting condition. No cure, nothing to be done. So of course, you know, you go and quit being a hedge fund manager and build a team and try to do it because egomaniacs, right?
10:54 Speaker 01 My wife is a applied behavioral analyst who just treats kids with autism. So I live that life.
11:02 Speaker 02 Exactly. So applied behavioral analysis uses variable rewards and a lot of the stuff you use in video games to rebuild words. So if you have a stroke or you have autism, you haven't basically said can, right? So the word can can mean I can, it can mean that can, it can mean the can can, lots of different things. But because there's so much noise in the brain, which is why you see a lot of kids and adults with autism not being able to handle large amounts of information. I have Asperger's myself and ADHD, they usually balance out, sometimes not quite. It's very difficult to pay attention and we all feel that, like, you know, when your leg is tapping, there's just too much going on. That's the fugue state because there's two transmitters in the brain, GABA, which calms you down
11:45 Speaker 02 and glutamate, which excites you. So when you pop a valium, the GABA levels go up. And imagine if your brain was always excited for one reason or another, you wouldn't be able to concentrate and build those connections to the word can, you know, that hidden layer of meaning for what can is. So, you know, that's why I repurpose drugs to adjust those levels. There's like 16 different things we identify that could potentially cause it. But because medicine treats everyone the same, it meant that one drug that made 34% of kids better made 28% of kids worse. So there's no cure and there's no treatment. So that's a different, bigger story and something again, this AI can help with. And why is that relevant? It's because what this AI does now is exactly the same. It uses giant honking supercomputers so that if you take a terabyte of text, it learns what
12:29 Speaker 02 the most important words in a sentence are. And that's what GPT-3 was from OpenAI. It took a thousand gigabytes of text and it learnt how to write in any style and any extension. So it played the most important part of any sentence. So you give it a sentence like Legolas in Gimli, that's all you give it and it'll write an entire scene in the style of Tolkien that's never been seen before. You know, you give it a sentence and you say make it happier, it understands what happier is. And this was called a latent space of meaning because it takes, uh, that took, uh, one terabyte of information, so a thousand gigabytes and GPT three is about forty gigabytes in size, but it can recreate any style of writing and it can
13:11 Speaker 02 understand any style of writing, which is again a bit crazy.
13:16 Speaker 01 I mean, it's, it's staggering. We'll, we'll come into the societal impacts and what this all means in a bit, but just want to get through the technology. It is astonishing when you see GPT three because basically it writes authentic pieces that are not stolen from other things, much like the deep, um, deep mind did with Go. It wasn't stealing from old games. It learnt how to do it. Is that right?
13:41 Speaker 02 It learnt, it learnt principles, it learnt styles, it learnt what's called a latent space, the hidden layers of meanings between different types of words. So like, um, we just released the most advanced image version of that, stable diffusion. Stable diffusion took a hundred thousand gigabytes, so a hundred terabytes of images, and we made a two gigabyte file that can do any style and any image of any type.
14:09 Speaker 01 So this is like DALL-E, right?
14:12 Speaker 02 It's DALL-E on steroids. So.
14:15 Speaker 01 So explain DALL-E and then we'll come into what you're doing because, you know, again, I just want to get people up the knowledge graph because DALL-E was the next big one that I stopped in my tracks. I went, holy shit, this is amazing.
14:27 Speaker 02 Yeah. So, so GPT three came out in 2020 and then we released the open source version of that GPT Neo, which I'll come back to. Last year, there was a breakthrough in image generation. So you can generate text and everyone's like, image is too difficult, right? Because a picture paints a thousand words, like surely it should be a thousand times more complex. Turns out it wasn't. So you had something called clip, which basically you took any image and it'd be able to classify it. So it'd be able to say that's a row and that's a red sofa and all these things, understand loads of concepts because they compress that down. Then you had a generative model that took words to images. So you had two models, a word to image model and image to word model. And at the start of last year, we figured out how to bounce them off each other. So generate an image and then check if that image was the same as the prompt and
15:12 Speaker 02 it'd go back and forth, back and forth, back and forth. And after 10 minutes, you generate a simulacrum of an image from the word input. So words go in, images come out. We're like, oh crap, it looks a bit mushy, but that's insane. It can generate an image of anything, but it isn't high quality. Then what happened is that OpenAI and us accelerated that whole space because we thought it was freaking cool and put lots of research into it until you got to the point now whereby you can put in Robert De Niro's Gandalf and one second later, you get an image output. So the first output of that was, the mainstream was Dali from OpenAI. That was in April of this year. So you can say a cybercore goth girl overlooking
15:55 Speaker 02 near Tokyo and boom, eight seconds later, it's generated.
16:00 Speaker 01 And this is a totally unique creative image that did not exist in the world before, right?
16:05 Speaker 02 It did not exist in the world before. All the data that went into that, so it was about 600 million images, it can't recreate any of those images. Instead, it's learned the principles of that. So again, it's the principle based analysis, which is insane, right? Because it's like, again, it's the heuristic stuff that we do all the time. And it can combine different concepts. So you can say a Van Gogh by Banksy, and it will do a Van Gogh by Banksy or kind of do Scream. And now there's more and more technologies that have emerged from that. So OpenAI kind of announced the closed beta of that. And then my company Stability AI created a version that was 30 times faster and more powerful.
16:48 Speaker 01 So tell me about Stability AI. When did it start? Who's involved? What are you guys doing?
16:55 Speaker 02 Yeah, so it started in 2020. Technically, it started in 2019 with the education work, but I can talk about that later. But 2020 is when it was formed officially to do the COVID work. So it's lead the UN COVID initiative, launched at Stanford, et cetera.
17:08 Speaker 01 So you founded this?
17:10 Speaker 02 I founded it, yeah.
17:13 Speaker 02 Because someone needed to build something to do AI as a public good. So I didn't exactly know what the business model was or anything like that then. I just thought someone needs to build this stuff because people are dying effectively. And then we had the other projects. Actually, I'll mention it. So my co-founder and I, so Joe runs the charitable arm. We have took the Global X Prize for Learning. So that was a fifteen million dollar prize funded by Elon Musk and Tony Robbins for the first app that could teach literacy and numeracy in eighteen months without internet. And we've been deploying it in refugee camps and low-income areas around the world. To the point where now we are teaching kids literacy and numeracy in thirteen months on one hour a day in a refugee camp in Malawi. And we just got the remit to educate every child in Malawi, three point nine million kids,
17:56 Speaker 02 with a completely open book of hardware, software deployment and curriculum. So we're going to invite the world to say, let's build an open source education system where the AI teaches the kids and the kids teach the AI. That's the type of stuff that we wanted to do because so much of the world's issues are coordination problems. So I figured out how to bring in the multilaterals, the locals and others, like this, and cool mechanism design. And so let's do education, let's do healthcare. But then as we were doing this and we talked to a lot of the big private companies, they promised a lot and they didn't deliver. And this caught me thinking about the future. So some of the people on this call will be familiar with the Chinese social credit score system and the use of AI in China to identify UGears and others and everyone's got a rating.
18:38 Speaker 02 You hang out with someone with too low a rating, your rating starts to go down. Eight million people in China can't use planes or trains.
18:44 Speaker 01 It's behavioral economics, right, done by AI.
18:48 Speaker 02 It's gamified life. Life has become a video game.
18:51 Speaker 01 That's right.
18:52 Speaker 02 And AI is the extra system on top. But now you think about it, the entire web two, as it might were, was basically Facebook and Google. It was AI.
19:02 Speaker 02 Big data models. And now you've moved to the big model era where to build these models, you need to have the best data, amazingly smart people and fricking supercomputers. By fricking supercomputers, I mean like supercomputers that are multiple times faster than NASA's supercomputers, because that's what compresses this data down into knowledge. So information to knowledge almost, because knowledge is when you extract the principles from information, right? And then wisdom is when you apply it to today. So we have this sudden knowledge extraction system. So I looked at that and I was like, the only people that can build this are Nvidia, Meta, Google, slash DeepMind, Microsoft, slash OpenAI.
19:40 Speaker 02 And how are they aligned? What does it look like? Because this is super powerful technology that can be used to target us ads better than anything else. Or it can be used to change our mind because it's convincing. You see G P T three writing. You can't tell it's not a human.
19:53 Speaker 02 Right? You look at the audio version, you can't listen, say it's not human like. My sister in law, Xena, just sold her company to Spotify, Senantik. Fully emotionally real A I voices using this technology. So she did the technology for all of Blizzard and everything.
20:09 Speaker 01 I've just seen somebody who's now got A I for creating podcasts. So you can have J Joe Rogan interviewing Steve Jobs. Neither of these exists and the interview never existed, but the A I does it.
20:22 Speaker 02 And it sounds just like them, right? And this is the thing, it's got that human level. So like, um, Senantik also did Val Kilmer's voice. For his documentary and Top Gun, because he lost it completely. And it's completely emotionally real. So it's getting to that period of emotional realness and human realness. Which means then you think about the future and you're like. This is one of the most powerful technologies ever because it's approaching humanity. Not in a generalized sense where the A I can do everything in its Skynet. But as a tool for targeting and convincing and manipulation. So what happens if only big tech companies do this? Because they went from being open to being private, even OpenAI. That got a billion dollars from Musk and others, and then a billion from Microsoft stopped releasing their code. Because they said it was too dangerous for people to have this.
21:07 Speaker 02 You know, so they turned closed and they stopped sharing. And you're like, as this is the most powerful thing, if you fast forward five, 10 years. Does it make sense that a private company will have a monopoly or a series of private companies have a monopoly on this? Then it will reflect basically Palo Alto norms. And it'll mean those companies are more powerful than any government in earth. And I was like, that's wrong. I was like, there needs to be an alternative that is open. You know, kind of A I for the people by the people we call it. Um, because not only is that morally right, because it's unethical to keep powerful technology. That can make people's lives better from the people that could benefit from it the most. But it's actually a better model.
21:49 Speaker 02 Because you have open innovation. This is why Linux became the servers of choice and the most secure servers. Windows has all these holes in, Linux holes get patched instantly, right? This is where you've got databases kind of taking over because nothing can beat human creativity when that happened. I look at Web three and Kripta and I'm like, it nearly got there and then it kind of got hijacked because it was part of the puzzle, but it wasn't the whole puzzle. Because when I went to talk to a lot of Web three people, I'm like, where's the AI in Web three? They were like, well, we'll get to it. But I was like, the whole of Web two was AI. How are you gonna do Web three without AI? And when they gave examples like Aleph here and other things, you know, with the talking NFTs, the technology they used was our technology that
22:30 Speaker 02 we released that used GPT near, which I thought was quite interesting. So then I realized that actually a bigger thing here and a bigger reason to do this. But how do you compete with OpenAI and DeepMind and Microsoft and Google with their billions of dollars of budget, which was an interesting one. How do you do it? Right? Do you go and raise $2 billion? No, go do it another way.
22:55 Speaker 01 So what did you do? This is fascinating.
22:58 Speaker 02 I built community. So I built community. So in 2020, a lot of people were just like, you know, COVIDed and in quarantine. So we built up a Luther AI, which the base principle was let's create an open source version of GPT three, you know, where it's like.
23:14 Speaker 01 Did you create it or did you create it or clone GPT three and then allowed to build on top? How do you start this?
23:21 Speaker 02 Oh, so, so like there were five initial creators and then I came like a month after, along with some of the second wave and said, let's accelerate this up. So they looked at the model. So the data wasn't open. So we had to create our own data set. And so we had to crawl the internet and create a terabyte of data. The compute wasn't available. So we got a grant from Google, who provided the compute initially. And then the expertise, it was all like PhD students or self taught programmers and others were like, how was this work? Because they released an academic paper, but they didn't give any code. So we built the code from scratch. And then created a model that was seventy five percent as good, but it was available. So the models from a Luther AI, which kind of stability runs now basically via the top level.
24:06 Speaker 02 People, but we're going to spin it off into its own independent charity for a number of reasons.
24:12 Speaker 02 They've been downloaded twenty five million times now by developers. So anytime you see a chatbot or something like that, that's getting to human levels, it's probably going to be that. I was just released open source to the world. It's eighty percent as good as G P T three, but it doesn't matter because you can customize it and you can extend it. And you can run it on your own hardware. So it's a twenty billion parameter model versus one hundred and seventy five billion parameter models, eight times smaller. But you know, it's just like the big steel mills and the little ones. If you look at the kind of Clayton Christensen thing, small can out compete big. And then I realized that there's a talent arbitrage here whereby there's a lot of people that want to build open source. But there's no supercomputer in academia. So when I said supercomputers needed for this, I meant it.
24:53 Speaker 02 Like the amount of supercomputer, the amount of breakthroughs in this foundational model research. Twenty years ago, a hundred percent came from academia. Ten years ago, seventy five percent came from academia. Last year, zero percent came from academia.
25:11 Speaker 02 Because there was this massive ramp up in compute capabilities. But academia couldn't access it, only private companies could. So if you wanted to have a breakthrough or work in cutting edge research, you had three options. Do a startup, startups suck. So just put that to the side, especially for academics, right? Number two, you go and work for big tech and then you get fifty nine page N D As. And you might do cool stuff like we've seen text to video from the Google team. That'll never be released by Google. Because of ethical fears, which we can discuss. And the final option is that, you know, you go and work for some of these independent labs like OpenAI and others that were meant to be independent and then became less independent over time. Otherwise you stay in academia, like I would like to do these, but I can't. So I saw that the core choke point was compute. A community was forming for talent.
25:56 Speaker 02 And then we need to be highly structured about data. So what I did was I found how big a supercomputer can I build if I put all my money into it? And I've become very persuasive. And then I built the tenth fastest public supercomputer in the world in four months.
26:12 Speaker 02 And it turns out you can build a very large supercomputer. So Ezra one, which is our supercomputer, which all my money went into is about eight times faster than the fastest supercomputer in the UK and about seven times faster than all of NASA's supercomputers put together. And that's the level that you need to basically create these models. It's like the entry level.
26:33 Speaker 01 But that need for compute keeps going up. Can you use distributed computing power? Or is it still not fast enough?
26:41 Speaker 02 So you can do it once you've done the initial model. So what happens is that you take, in this case, we took a hundred thousand gigabytes of data. So we made the biggest image dataset in the world. Previously, the largest was a hundred million, then we made four hundred million, then we made five point six billion, a two hundred and fifty terabyte image label pair dataset. Then we took two billion of that, which is the high quality images, and then we crunched them back and forth on this. So what happens is after you crunch them back and forth, then you've got a file that can be adapted and trained. It's like, basically, it's gone through primary school of learning. And then you can teach it more advanced concepts. So like, some people are taking it and training on Japanese concepts. Some people are taking it and training it on anime, cat, girl, waifus, or whatever, you know?
27:25 Speaker 02 You can add that specialization once it's achieved kind of the high school level. So there's a big compute and a little compute.
27:31 Speaker 01 Does the network then learn from all of these people working on the network itself? So I
27:36 Speaker 02 No.
27:37 Speaker 01 You've got, because you're distributed and you've got lots of people doing different things, does that increase the knowledge base or not?
27:44 Speaker 02 It increases the large base of the community, but you have to understand this model. It's not a distributed model where it's like lots of little brain cells out there. It's a two gigabyte file that can recreate and create any image in any style. So you took a hundred thousand gigabytes and created two. So sometimes I think I'm on that Silicon Valley show, you know, from HBO, and we're Pied Piper, and I think, crap, am I Erlich Bachmann or Ross Haneman? You know? What is my role in here? It's one of those kind of guys, right? It's the most advanced technology the world's ever seen for compression because the entirety of humanity is basically about communication and compression, right? We're compression machines, that's all we do.
28:27 Speaker 02 So people are listening to this podcast now and they're literally listening to yours and mine common context and the compression of all the knowledge that we've had over the years. And they're learning something new right now. It's what we do. And hopefully it's valuable because they'll see that holy crap, this new technology wave is coming. But if you can do it to images, you can do it to anything. So actually, how we had the breakthrough is we took a language model and image model and we fused them together. And somehow it learned the different concepts.
28:53 Speaker 02 Then what we're seeing is that when you have different models of different types, so someone takes a two gigabyte file. You take it and you put it all your images in and I take it and I put all my images in. We can actually fuse them together and create another two gigabyte file that knows both of them without going up in size very much. That's a bit insane. In fact, OpenAI did this with a model called GATO recently. So what happened is that there were hundreds of different models they built that were quite big. That were like robotics and playing chess and all these other things. They said, what happens if we combine them all together and see what happens to these latent spaces, the hidden meanings of understanding, just like the neurons in our brain?
29:31 Speaker 02 It created a 1.6 billion parameter file that can open doors and play chess and play Starcraft and all sorts of other things. And they were like, what?
29:42 Speaker 02 And this is the thing, like you've got a breakthrough in intelligence. But then you've broken through to the point where, this is the important part. Stable diffusion is the first model that's small enough, fast enough and cheap enough to go anywhere. So you can run it on your MacBook M1. You can make Robert De Niro by Banksy and Renoir in Guadalajara in a snowstorm without internet access. And that two gigabyte file will reproduce that faithfully each time.
30:11 Speaker 01 Without internet access.
30:14 Speaker 02 Yes.
30:16 Speaker 02 It is a two gigabyte file that has compressed the knowledge of the internet in images.
30:21 Speaker 01 I can't, I can't quite get my head around this. So where do you think what you're working on goes? What is the image thing? Where is this going towards? Because right now it's kind of interesting. Yes, we're going to see creative industries, marketing, others use this quite quickly and effectively. But where is this in your head? Where's your mental model? What you're actually working on here? Because you're not working on creating nice images. You're working on something bigger. What is that?
30:52 Speaker 02 It's the intelligent internet. Intelligence is about to be pushed out to the edge. So every person, country, country and culture has their own models that are constantly updated. So you'll have your own model of all your knowledge across modalities. And you can run it on your local hardware without being accessed to the internet. So the internet right now is centralized and all is running on Google servers and guiding you. That gets pushed out to the edge. Because finally you have this compression technology and, you know, it's interesting because there's a bunch of structural things here. So Apple is the most interesting thing with regards to this. Apple will be the biggest AI company in the world in two years. Why? I'm listening to this on a MacBook M1 right now. Sixteen point eight percent of the chipset is a neural engine that's not used.
31:37 Speaker 02 That is designed for exactly this type of model. And it's proliferating, right? Everyone's moving to M1 MacBooks. Everyone's moving to A16 Bionics that also have neural engines and they're not used. Are Apple stupid? No. They have an entire teams building these types of models for their augmented reality and beyond. So you're going to move from Siri one to Siri five. The Siri's a bit crap. It won't be soon. And so as that technology proliferates and then people take what we're building. And you create a hundred thousand, a million new developers in this space. The possibilities are endless. You've seen hundreds of things built on stable diffusion. But like I said, I don't just want images. I want audio. I want video. I want to have text. I want knowledge. I want my databases combined with it. I want my type one and type two brains.
32:21 Speaker 02 And then you have an AI that either manipulates you or it works for you. And my thing is it's open infrastructure that should work for every individual. And this is also why I've got a very different approach to that. So most people's approach would be B to B. Go to the big companies and sell them API access. Instead, I'm forward deploying engineers into the biggest brands in the world. Getting them to invest in my next round of financing. And I'm also doing that for countries. I'm going to India and saying, what's the probability this technology will be used by everyone in India in 10 to 20 years? A hundred percent or 95%. So I'm saying, I will build it for you next year and getting all the biggest Indian conglomerates and others together and taking the smartest engineers out of Silicon Valley back to India.
33:02 Speaker 02 Like one of my fun ones is I have a JV with Eros, which is the Netflix of India. So 200 million daily active users, all the Bollywood content, I have an exclusive on it. And a revenue share. So what's the probability that Bollywood content will be interactive through these models? Of course it will be. So I'm locking down all the content.
33:22 Speaker 01 Netflix are really working on interactive content, actually.
33:26 Speaker 02 Yeah, but not like this. They don't have this expertise because right now, you know how many people can build these models?
33:32 Speaker 02 Forty.
33:35 Speaker 02 In five years, how many people will be able to build these models? Four million.
33:40 Speaker 02 So there's some really interesting arbitrages in the market right now. And my take is that if I push hard enough and I out compete the big guys, which I am, like my team's 100 now. Um, you know, and I'm out competing these big guys by building markets, products, markets, uh, products to the market quicker.
33:59 Speaker 02 Then I force everyone to go open source because it's no longer a differentiating factor. And then that changes it from a closed panopticon for our kids, you know, and ourselves to something different. I'm not sure what it is, but I think it's more beneficial.
34:12 Speaker 01 Okay, so let's talk about what it could be. It's clearly fucking terrifying. Because either open source or in private hands. If it's open source, it's like a virus. It'll go where it goes and people will use it how they use it. And there's no stopping it. Okay, fine.
34:35 Speaker 01 How does society deal with the fact that we don't know who is a human and who is not? And how we're being manipulated, who not? I wrote a thread on this the other day about digital ID being one of these key things, but it's this whole element of deep fake of what is real in this world? How do you deal with this?
34:58 Speaker 01 Because it's going to shatter society.
35:01 Speaker 02 It will adjust society, definitely. And so, part of my reason for releasing the model, you know, during the COVID thing, I just had a lot of conversations about third immunity, right? Which was bad for that. I believe it's good for this. Releasing this model out there, it's going to be everywhere in a year. This is a one-to-one billion person moment, right? Right now, a few million people.
35:23 Speaker 01 This is exponential. This is exponential without question.
35:27 Speaker 02 It's exponential without question, but this is the really interesting thing. Right now, very few people know that you can create anything in a second for one cent and less than a cent soon, right? Soon, everyone will know about it because it'll be in all your favorite apps. So we're integrated into Canva right now. We're integrated into Photoshop. You can see big announcements coming out because we're partnering with everyone with interactive content. What does that mean? It means people suddenly start thinking about this thing. What is real? What is not real? And I said, identity is key. So the other part of Apple, I'm a bit bullish on Apple, is the fact that they've got the identity architecture down. And that allows them to have verified content creation. So one of the things we have with Adobe and others is we've been pushing content authority, basically, which is a small metadata pile that attaches to every piece of content and its hash. And
36:12 Speaker 02 if either of them are edited, it changes and it shows that it's not true. It's not a blockchain thing, but it allows you to verify content in a positive way, like it's
36:23 Speaker 01 Why is it not on blockchain?
36:26 Speaker 02 It's not on blockchain because it doesn't need to be on blockchain, that's why.
36:29 Speaker 01 Well, because the centralization of the knowledge, right? Because like Google owns all of the knowledge. Blockchain is the only way of not having one central power owning the knowledge.
36:40 Speaker 02 It's a, it is a way of not having one central power owning the knowledge and you can combine this with a blockchain. But you can think of it as super advanced metadata. So you know, when you have metadata in a file like EXIF, when you edit that, the file stays constant. Did this, they're both immutable effectively. So you can tell if one of the others been tampered. So if it says that it's from Raul, it's from Raul effectively. But there's no blockchain lookup for that. You can make it stronger with a blockchain. But you know, it's just a standard that's open source that we're pushing effectively. So it's kind of a positive one. It's not a negative one. Obviously there's ways around that and things like that. But it'll get people thinking about these things. How does identity work in a new age? Cause so much of what we do is identity, right?
37:23 Speaker 02 Most of finance is basically securitization, which is identity. Most of kind of, the whole of blockchain is identity. It's all about identity exchange. So the thing is, how can we build better systems for identity? Cause that's the final part, information, this massive information. And also stable diffusion, as we released it was a snapshot of the internet. It's biased, it's racist, it's everything. If you enter the prompts in wrong.
37:47 Speaker 01 Yeah, I remember you showing so many images of the racial bias of the algorithms.
37:53 Speaker 02 So my thing was like, if it's just one company that controls it, like what OpenAI did with DALL-E. So DALL-E is a control system. You don't access the code images or anything. To make it less racist, whenever a gender neutral word is put like sumo wrestler, it would randomly add a gender and a race. So you type in sumo wrestler and you've got Indian female sumo wrestler, tiny little Indian lady. Sumo wrestling and stuff like that. Technically, yeah, I mean, that is one way. But instead, my preference is for every single company, country and culture to have their own models and be able to create their own. And then like I said, then it becomes acknowledged what it looks like, but then you can set standards. You can be the standard that people build around and you can incorporate authentication in that standard.
38:39 Speaker 02 Because the other way that I talk about these things, it's a generative search engine. Cause now they have stable diffusion and in six months, it will be perfect photorealistic, 12 months maximum. Do you need Google image search anymore? Cause what's your job to be done when you do Google image search? It's to have a picture of a certain type.
38:59 Speaker 02 Now a little fricking two gigabyte file on your local computer or something that you pay a fraction of a penny for online can create any image you can imagine. And you can edit it with your words, basically.
39:10 Speaker 02 What is that if not a search engine of a type? It searches for a concept and it turns it into an image.
39:16 Speaker 01 It's a creative search engine. It creates by, as you say, you just put in a word or a vocal command and say, I want this.
39:25 Speaker 01 And it creates.
39:25 Speaker 02 But then, but then you can also say, I want a presentation about cicada migrations affecting sesame seed prices. I watch too much Silicon Valley here. And in a couple of years, it will create a beautiful presentation for you. And then you can say, I want to be happier or sadder, or I want it to be more impactful. And it will adjust the presentation dynamically for you. No more PowerPoint.
39:51 Speaker 01 But it's going to do video as well, right? We don't need to do this. I could just say, hey, have a conversation between Raul and Ahmed about AI. And here's the narrative arc and it'll just do it.
40:01 Speaker 02 Pretty much, yeah. Yep, just like it does in Steve Jobs, but in 3D video. So like, I don't know if I should be talking about this, but I'm going to talk about it anyway. Well, so what if I get in trouble?
40:12 Speaker 02 There are two, well, I heard you. There are two boxing legends families who are giving us all of the motion capture data and media from their lives. And we're building up
40:26 Speaker 02 an analysis of how that works so that we can basically say, who won between X and Y in a fight in their prime, run on the supercomputer where nobody knows the outcome in full resolution.
40:36 Speaker 01 When they never fought with each other because they were of different generations.
40:40 Speaker 02 Yes.
40:42 Speaker 02 So you can pick two, you can pick two different people. I can't say who they are, but you know, those types of things that you've always postulated. We can do that now.
40:51 Speaker 01 And obviously this is going to completely change music as well, right?
40:56 Speaker 02 Yeah, so we've released our music models already, Dance Diffusion, where we did a slightly different thing, we want to take a snapshot of the internet. We made it so everyone can create, take their back catalogue, incorporate it into a kindergarten level of knowledge. We decided not to go to primary school for the model. And then you can query that to create your own style in anything. So we call it Dance Diffusion, and we're teaming up with the top EDM DJs and others in the world, so that every musician will have their own models that can generate music in their style. And then, if they give permission, we're going to mash them all together, and you've got a generative Spotify, effectively,
41:35 Speaker 02 which then, basically, as people prompt it, there's an attribution system where it pays them. And that will just be like 100 gigabytes as a file that'll capture the music of the world, which again is insane. This compression thing, you know, it's actually like, you know, you read so much, right, Raul? And then you write your notes, and then you write your investment thesis. An investment thesis is compression of knowledge. A vaccine schedule was compression of knowledge. We finally figured out how to make computers compress information into knowledge.
42:07 Speaker 01 I mean, this is, I knew this was going to be an interesting conversation, and I've been looking at this for a while, but frankly, I'm fucking staggered at how fast this is moving and what you're doing. I mean, it is, I mean, this is truly exponential, as you say. Within two or three years, this is multiple billion people, and a complete shift on a humanity level of how information is processed and delivered.
42:37 Speaker 02 And that is society. So basically, it's a shift in society. And so my thing was to go into emerging markets and others and give us technology to them. Otherwise, they'll be left behind, because, you know, like Dali 2, you can't use if you're in Ukrainian, and you can't use it for anything Ukrainian-related. And there's no appealing that or anything like that, you know? But it means Ukrainians are left behind, whereas everyone else can become digitally augmented creatives that can generate any image in eight seconds. So this is also why at Stability, we do everything. We do code models, image models, language models, audio models, video models, protein folding models and others. And so you're going to see leaps across all these areas in parallel.
43:16 Speaker 01 I've argued for a very long time, in fact, I was writing articles in GMI about this. Ten years ago, that AI plus big data equals medical breakthroughs. Because what the AI can do with that data is something that humans can't do in the standard scientific method of hypothesis testing. It's just too slow and too cumbersome.
43:38 Speaker 02 You can get it to generate null hypotheses, which humans don't like to do, you know? You can use this to transform so many things once it gets productized. Because the other thing, most of this stuff was stuck in labs. It wasn't productized. So if you look at what's time and stable diffusion, people have built hundreds of products on this ecosystem already. Like make your own sneakers, put yourself into any movie. You know, they've done architecture, they've done 3D worlds. Like the Cambrian explosion of talent is similar to what I've seen at the start of early days of Web three. Except for there's no reason to bootstrap economic incentives when you're creating value. A lot of value in the world, to be honest, is the entropy of information. How much value in the world is taken by taking unstructured data and making it structured?
44:22 Speaker 02 That's exactly what this technology enables anyone to do. And so the future must be that we have AI that works with us to structure the world around us to enable us to achieve more. That's kind of my theory here. And this is again, why, like I said, it needs to be this person to invest.
44:37 Speaker 01 So you would err towards augmented humanity as opposed to a singularity? Or is it augmented and then a singularity? Where does this go?
44:49 Speaker 02 So look, I find the whole AGI thing largely distasteful for a variety of reasons. I think it misaligns a lot of incentives. And I think the current way that a lot of these big labs are going, which is that you take the data and you create gigantic models that are only usable by gigantic machines will probably kill us all if it creates a singularity. Because the incentive alignment of these groups is to serve us ads and manipulate us effectively. And it's trained on largely Western data sets. You know what happens when you, like, there was this thing, Lisa, the chatbot from, was it Lisa? No, Tay, the chatbot from Microsoft. You remember that?
45:25 Speaker 01 So it was a chatbot they put on Twitter.
45:26 Speaker 01 Stop it because it became abusive.
45:29 Speaker 02 If it became a Nazi, so it talked to people and learned from them. Within a day and a half, it turned him into a Nazi, you know, like if I train on the internet, I'm gonna create a really fucked up AI, to put it quite honestly, right? It's not gonna be an AI that loves, it's gonna be an AI that probably freaking kills us all. Whereas if everyone's got their own AIs, and this came through your thing, like, does a network learn? If we take the models of all of humanity and all the cultures and all the different ethical views, and we combine those models, and all these AIs are designed to augment humanity rather than manipulate them, I think that's a far more positive potential. I don't think it's sufficient, I think a lot of work needs to be done around this. AI is dangerous, but it's inevitable. And the way that it's being controlled by private corporations, again, you know, I'm
46:14 Speaker 02 a capitalist, you know, I'm gonna build a trillion dollar company to help 10 billion people or whatever. But I think the current mechanism is wrong. And I think, again, there's a lot of things around the singularity thing that are misaligned. If we do build singularity, I would like it to help us all. But in the meantime, let's just help people. Like, let's make people more creative, let's make people able to access information.
46:35 Speaker 01 Yeah, but there is unintended consequences, you get that, right? There are very unintended consequences. We have no idea what the probabilistic outcome is of this.
46:44 Speaker 02 This is decision making under uncertainty, not risk, right? So this is why the default for all of the corporations is this technology should not be released to anyone, because we're the only ones responsible for this technology.
46:56 Speaker 01 And that in itself doesn't stand up to scrutiny. And that doesn't stand up to scrutiny. Both, there's different outcomes from different methodology, giving it to everybody, who knows what foreign governments do with, you know, there's a lot here. The other thing-
47:09 Speaker 02 Well, I can tell you on that one, foreign governments already have access to this technology. So Russia and other, China and other places ramped up their supercomputers massively. So they already have access to far more advanced versions of this technology than the one that's being made widely available. Which is one of the reasons we need to build a higher immunity ahead of the next election cycle.
47:27 Speaker 01 So you think that the, the, the, let's use the Russian example, a Russian online misinformation campaign is now being run by AI, which is why it's pretty much unstoppable?
47:40 Speaker 01 The best.
47:40 Speaker 02 It's probable if you if you build better AI. I mean, to be honest, you like Twitter and things like that, like, come on, man, you can do better than that. Like, you see all the checkmark hijacking and things like that. You can tell what an AI is if you've got sufficiently advanced AI on the other side. But again, people learn to not trust everything they see until we introduce the new structure. Like, we right now need to have a verification protocol for information. Our social media and our information systems aren't adequate to the task. Uh, the internet is, uh, no, I think there's Max Clifford from EF that said this. It's an intelligence amplifier, but only for the few. So what happened is the internet took it, compressed it down, and then some people just went way out there, and they have disproportionate impacts on everyone, and most people are left behind.
48:25 Speaker 02 They have the audiences, they set the narrative, and the narrative that basically is most appealing to the mammalian part of our brain is one of divisiveness, which is why the middle has disappeared. And that's why it can tap into these things.
48:38 Speaker 01 That's right, exactly right. Because, you know, the easiest way to manipulate humans is emotion, and the strongest emotions are the polarizing emotions.
48:46 Speaker 02 Whereas with this technology, you can say, I want this article written from the perspective of a tea party conservative versus a libertarian, and I will automatically change that. It's a universal translation engine as well. Again, which is crazy.
49:02 Speaker 01 Clearly, clearly, yeah, yeah, without question. So, okay. Once we've got this technology, we've got another super trend that's happening at the same time. Obviously, there's a few super trends all happening. I call them the exponential age. One of them is compute power, all of this, right? So they all go together. The other one is robotics. So now, how close are we to creating sentient robots? Where is sentient in this process?
49:29 Speaker 02 What's the definition of sentience, Rob?
49:31 Speaker 01 I don't know. That's the question. We don't know, right?
49:34 Speaker 02 Nobody knows what sentience is. There is no commonly defined thing. Um, in terms of like, the things you see in the movies. Nobody knows, because this is the thing. If you ask anyone how far away from singularity or intelligence, um, people will say at most, at minimum, eighteen months, right? Um, what is it? Why? Because what possible information could you have to say that a human-level intelligence is less than twelve months away?
50:05 Speaker 02 I'm not sure and nobody's been able to tell me an answer. So, you know, this is a case of we don't know, but what we do know is that AI is getting better than humans at certain things. So like OpenAI just released, maybe they were pressured by someone, the open source version of a software called Whisper, which is a dynamic transcription engine. You can speak in five different languages in the same sentence and it will transcribe it perfectly into English. It's actually got above human level transcription quality. And so stable diffusion is above human level image generation quality. GPT three is above human level writing quality. When we combine those all together, you will get an illusion of sentience.
50:46 Speaker 02 Because you can't tell it's not a human, maybe, but is it really sentient and has agency? I don't know. And for the robotics thing, I think that Tesla's got some much bigger plans than they're letting on in regards to that. And you're going to see some big speed ups.
50:59 Speaker 01 Yeah, I again, I think people. People are underestimating the speed of what is happening. And I think you've confirmed it to me that the speed of which all of this is happening is is ridiculous.
51:14 Speaker 02 It's it's a true exponential. This is also actually when you look at AI research papers, it's a doubling every twenty four months. On that. So it actually is an exponential thing on AI research papers when you plot it on a log graph is a straight line.
51:30 Speaker 02 So as this goes, it will continue going exponential. And like I said, exponentials are a hell of a thing. We are not equipped to handle them.
51:37 Speaker 01 No, and we've got too many get well, too many. We're going to have the fastest pace of change humanity's ever seen because of technology. Because there's so many of these all happening at the same time. It's it's a lot.
51:50 Speaker 02 Even as everything else breaks down, all our systems are at the edge. That's why I called it stability. Because I was like, we need to reform education, healthcare and all these things quick. Because we're going to see what we're having to. This is not normal, what we're seeing today. You know, like they might say, oh, interest rates are seven percent. What worse that could happen? Come on, guys, like, these are symptoms, not kind of causes. The reality is our systems are outdated and need to be improved. So what happened with crypto is they tried to grow a whole new system and then there's this system. And the interfaces where all the money is made and lost. Whereas this technology, because it can take unstructured data to structured data, can sit in our systems and extend them.
52:30 Speaker 02 They can disrupt them or it can extend them. So it's really interesting. I think, again, the pace of change will be ridiculous.
52:37 Speaker 01 So how the hell would somebody like Real Vision deal with, I mean, every company on the world in the world has to change their business models yet again. People are changing to Web three because that's another new business model. But here we've got the application of technology at scale in a totally disruptive way. How the hell do we all deal with it? How do we even get ahead?
52:58 Speaker 02 So the lovely thing is that, you know, when I go into the biggest media companies in the world right now, they've already prototyped the technology internally because we made it so they could. This is a new internet that's coming. And so the reason my self stability was to be the layer one for next generation AI. So my aim is to build the biggest company in the world that puts this technology to everyone so they can build their own models. Or if they want the white glove service, they come to us and we forward deploy engineers to take all the content in the world to make a living and interactive. But you've just kind of got to get into this. And again, you know how, remember how it was when you sent the first Bitcoin? You know, it's magic. Sufficiently advanced technology is magic. This is the same. When you create your first image, like if you do something boring, it's gonna be like whatever, right? It's style transfer.
53:42 Speaker 02 But when you create your first original image, combining different styles and concepts and the image in your head becomes reality. You're like, this is new, this is different, this is crazy. It's the biggest thing of all time and it's not just for images. Images are the breakthrough wedge.
53:59 Speaker 01 Yeah.
53:59 Speaker 02 Because it's gone from 20% to 80% language. GPT three was an 80% to 90% moment. This is 99% of the world don't believe they can create, they can suddenly create anything. You've got to just get in there and you've got to basically find where it is because what happens is that Pat and Christian had this wonderful description of infrastructure as the most efficient means by which a society stores and distributes value. The value landscape in content, in enterprise, in a lot of things, in information is about to be upended. And we don't know what the new landscape is. And so you've got to be in there to understand that.
54:33 Speaker 01 What role do humans play?
54:36 Speaker 01 You know, obviously, we've got a shrinking world population over time. So I think of robots and AI as demographics. And I wrote a few articles about it. So these are new demographics, but they can proliferate very far. So does that increase GDP growth? And humans need to move towards some sort of universal basic income or different roles? How are you thinking that?
54:58 Speaker 02 I think that we need to move towards universal basic income. Like, you know, this is why I'm running this out in emerging markets. The education tablets create the best data sets to feed models for every nation. So I'm going to make it so that Malawi adds percentage points to its GDP. Ethiopia adds plenty of points to their GDP. India is the number one market for this because they have Atar, India stack 5G and all capital needed. So India is one of our biggest markets by far. And literally, we will accelerate Indian GDP and the Indians will then have a better system than we have in the West. Like already 13 months to literacy and numeracy on one hour a day is better than most primary, you know, kindergartens, right? So think about what it's going to be when it's a self learning AI system. It's going to be crazy, but then they are equipped for that. And you need to have things like UBI.
55:43 Speaker 02 So one of the things we've done is in our country level subsidiaries, and we're doing just about all the countries in the world. 10% of the shares are reserved for the kids that use our education tablets. So they'll be getting shares as they kind of grow. And then we're trying to figure out how UBI looks like on the back of that, which is insane. Like, we shouldn't have to do that, but nobody else is thinking about the pace of change and scale of this, which is why.
56:06 Speaker 01 So are you going to structure this as a DAO or a foundation or something?
56:10 Speaker 02 I wanted to do a DAO of DAOs originally. And then I realized the infrastructure wasn't there. And then I realized, let's just IPO in every single country and create the next trillion dollar business. That's the best way to do this.
56:23 Speaker 01 By distributed. Distributed listing. Okay, interesting.
56:28 Speaker 02 Yeah, distributed listing. The Indian version of stability should be owned by the Indians for the Indians. We're the smartest Indians running it. And there's an opportunity here, you know, same for all these other countries. Maybe DAOs work in the future, maybe you can do airdrops. It's open, right? But this is a cool thing. This technology has an exponential.
56:50 Speaker 01 So.
56:51 Speaker 02 Who knows? Nobody knows what the answer is.
56:54 Speaker 01 Final question, because there's a lot to digest here. How are governments going to deal with this?
57:02 Speaker 02 So Wes, the European Union is trying to ban open source artificial intelligence now because they view the edges regulation. So authors of models will be liable for the use of the models. The US is uncertain right now. The UK is massively pro AI and so they've been upgrading it massively. Uh, in emerging markets, guess what? I'm going to the governments and telling them they can have this technology for free. They love me. So you'll have most of the world loving it. And then Europe is a big question mark. Uh, US is a small question mark. And I think they'll be very pro this because it adds GDP at a time when there is no abundance. Like, the question is, does it then collapse GDP afterwards?
57:45 Speaker 02 Parties don't think that far, honestly.
57:48 Speaker 01 No, when I look at it, you know, what is GDP? GDP is population growth plus productivity. This does both population growth and productivity. Cause it is population, right? Robots and AI are demographics and it increases productivity. So that should mean the pie increases. So per capita GDP rises. That's my working hypothesis until it replaces the humans eventually anyway.
58:15 Speaker 02 Exactly. And there's this other part, which is that basically the West and advanced economies have borrowed too much from their balance sheet from the future based on identity. Whereas you look at India and other places, they haven't done that. So when you incorporate this technology with government entity structures, you suddenly have massive credit creation like you've never seen before. Because information, which money can flow much nicer around India or fricking Ethiopia or anywhere else, especially because our tablets are standardized tablets that we're deploying at scale with a new type of bond whereby you only pay based on outcomes if you're a donor, which is measured through the tablets. So my thing is like infrastructure uplift for those countries. The West, I have no idea what to do with. Like, I'm just going to the companies, saying, give me your content and let's share in this upside.
58:58 Speaker 02 Um, but it's difficult to try and fix the UK or US or Europe. And there's enough smart people there who can do that.
59:04 Speaker 01 Amazing conversation. I've absolutely loved it. I'm terrified and excited at the same time, which is what I was hoping. I'll definitely get you back to talk more about this because we barely scratched the surface.
59:17 Speaker 02 Yeah, no problem. Like I said, we've kind of not talked about it publicly. I think this is one of the first times we're talking about actually the bigger plans. Like I think we're recording this now, next week we do our big launch event. But everyone needs to know about this, right? And everyone needs to participate because this should be a communal effort. Because we need to guide it together, right?
59:34 Speaker 01 Exactly right. Well, listen, best of luck with everything. Let's hope the unintended consequences are not as bad as could be. We'll only find out, but we're going down this path anyway, regardless. So you might as well give it to the masses and not to the few.
59:49 Speaker 02 Exactly. Cheers, Ralph.
59:52 Speaker 01 Thank you, my friend.
59:56 Speaker 01 There is so much to unpack for an interview like this. Firstly, it shows that the hypothesis that technological growth is going exponential and it is not going to stop regardless of what the central banks are doing or inflation's doing. It's kind of irrelevant when the speed of technological adoption is so rapid and so truly extraordinary and profound.
1:00:24 Speaker 01 The ways what this is changing humanity, both from education, medical sciences, science in general, the creative industries, the media industry, everybody. This is again happening on a scale like the internet, but maybe even more so. And we've got many of these technologies all overlapping. As I've talked about before, blockchain technology has been the fastest adoption of any technology the world has ever seen, and looks like AI may even exceed that. And then we've got robotics and space and internet of things and on and on and on.
1:01:04 Speaker 01 So you can either fear this stuff,
1:01:08 Speaker 01 or you can use it to your advantage and invest in your own future. And that's the route I take.
1:01:14 Speaker 01 There is unintended consequences. We just don't know what they are. And I think Ahmad's point about should it be given to the few? Or the many to deal with those unintended consequences? I think he's probably right that give it to everybody is probably the only way around this. Anyway, a truly extraordinary interview and I hope you enjoyed it.
Previous December's Stock Sell-Off Rages On Next The Kyle Bass Interviews - If We Don't Get it Right, the Consequences are Dire Page updated October 27, 2026 · transcript and summary are machine-generated