AI Power Wars | Emad Mostaque

Today on the show, we have the founder of Stability AI, Emad Mostaque. Emad recently left his company citing “You can’t beat centralized AI with more centralized AI” and decided to venture into the frontier of decentralized AI. We touch on the economical consequences of AI, why it needs to be open-source and distributed and the role of Crypto in decentralizing AI. ------ 📣 SPOTIFY PREMIUM RSS FEED | USE CODE: SPOTIFY24 https://bankless.cc/spotify-premium ------ BANKLESS SPONSOR TOOLS: 🐙KRAKEN | MOST-TRUSTED CRYPTO EXCHANGE https://k.xyz/bankless-pod-q2 ⁠ 🔗CELO | CEL2 COMING SOON https://bankless.cc/Celo 🏠 CASA | SECURE YOUR GENERATIONAL WEALTH https://bankless.cc/Casa 🛞MANTLE | MODULAR LAYER 2 NETWORK https://bankless.cc/Mantle 🗣️TOKU | CRYPTO EMPLOYMENT SOLUTION https://bankless.cc/toku ------ TIMESTAMPS 0:00 Intro 5:26 Navigating AI & Web3 10:52 AI Governance 18:03 Decentralizing AI 26:54 Balancing Coordination 33:52 Decentralized AI Infrastructure 40:10 Emad’s Projects 48:40 Decentralization Trade-Offs 58:53 Governance Structure 1:02:50 AI Regulation 1:06:36 Custom AI Models 1:12:21 AI Economical Consequences 1:17:11 AI x Crypto 1:20:47 Closing Thoughts ------ RESOURCES Emad Mostaque https://twitter.com/EMostaque ------ Not financial or tax advice. This channel is strictly educational and is not investment advice or a solicitation to buy or sell any assets or to make any financial decisions. This video is not tax advice. Talk to your accountant. Do your own research. Disclosure. From time-to-time I may add links in this newsletter to products I use. I may receive commission if you make a purchase through one of these links. Additionally, the Bankless writers hold crypto assets. See our investment disclosures here: https://www.bankless.com/disclosures

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Web3 has grown up a lot, but there are actually primitives now that are emerging that could be a substrate for collective decentralized distributed AI by the people for the people. Whereas on the centralized thing, again, it's a different paradigm, which is the all-powerful AGI. My thing is, you can't beat centralized AI with another centralized organization. I think what we should build for an AGI, this generalized intelligence, is the human collective intelligence, amplified human intelligence, something that uplifts us all and acts like a swarm. The alternative picture being presented here by DeepMind, OpenAI, and others is the machine god.

• 0:39David Hoffman?

  Welcome to Bankless, where we explore the frontier of internet money and internet finance. And today on Bankless, we explore the frontier of decentralized AI. Imad Mostak, the founder and former CEO of Sability AI, is on the show today. Imad recently left Sability AI citing "you can't beat centralized AI with more centralized AI," and announced his intention to work in the nebulous field of decentralized AI. Decentralized AI, what exactly is that? Why does Imad want to put an AI model into the hands of every individual with the compute power to run it? Is there a race between centralized AI versus decentralized AI, or do they each bring something to the table? Why are the interests of governments,

• 1:21David Hoffman?

  nations, and communities of the world aligned with the desires of decentralized AI? Why will this revolution be deflationary for developing countries but an accelerant for the developing ones? Bankless Nation, if you haven't gathered this by now, it's an AI episode today because it seems the world of decentralized crypto protocols and the rise of AI are converging faster than ever, and we at Bankless are just trying to keep up one podcast at a time. Ryan, the AI, is out and unplugged for this episode, hanging out with his AI wife and AI children, giving us a chance to catch up as he takes a much needed break from the internet and meme coins today. Let's go ahead and get right into this episode with Imad Mostak. But first, a moment to talk about some of these fantastic sponsors that make this show possible. Especially Kraken, our preferred place to trade crypto tokens. If you do not have an account with Kraken, consider clicking the links in the show notes to getting started with Kraken today.

• 2:06David Hoffman?

  If you want a crypto trading experience backed by world-class security and award-winning support teams, then head over to Kraken, one of the longest-standing and most-secure crypto platforms in the world. Kraken is on a journey to build a more accessible, inclusive, and fair financial system, making it simple and secure for everyone, everywhere, to trade crypto. Kraken's intuitive trading tools are designed to grow with you, empowering you to make your first or your hundredth trade in just a few clicks. And there's an award-winning client support team available 24/7 to help you along the way, along with a whole range of educational guides, articles, and videos. With products and features like Kraken Pro and Kraken NFT Marketplace and a seamless app to bring it all together, it's really the perfect place to get your complete crypto experience. So check out the simple, secure, and powerful way for everyone

• 2:50David Hoffman?

  to trade crypto, whether you're a complete beginner or a seasoned pro. Go to kraken.com/banklist to see what crypto can be. Not investment advice. Crypto trading involves risk of loss. Mantle, formerly known as BitDow, is the first DAO-led Web3 ecosystem, all built on top of Mantle's first core product, the Mantle Network, a brand-new, high-performance Ethereum Layer 2 built using the OP stack, but uses Eigenlayer's data availability solution instead of the expensive Ethereum Layer 1. Not only does this reduce Mantle Network's gas fees by 80%, but it also reduces gas fee volatility, providing a more stable foundation for Mantle's applications. The Mantle Treasury is one of the biggest DAO-owned treasuries, which is seeding an ecosystem of projects from all around the Web3 space for Mantle. Mantle already has sub-communities from around Web3 on-boarded,

• 3:34David Hoffman?

  like Game7 for Web3 gaming, and Bybit for TVL and liquidity and on-ramps. So if you want to build on the Mantle Network, Mantle is offering a grants program that provides milestone-based funding to promising projects that help expand, secure, and decentralize Mantle. If you want to get started working with the first DAO-led Layer 2 ecosystem, check out Mantle at mantle.xyz, and follow them on Twitter at @zeroxmantle. Bankless Nation, super excited to introduce you to Imad Mostak, who has a really interesting background, was once a hedge fund manager, but moved to finance to become the CEO and founder of Stability AI back in 2020, the company behind Stable Diffusion, which is a text-to-image machine learning model that rose to prominence alongside ChatGPT. I met Imad about a year ago during the AI Crypto Week.

• 4:16David Hoffman?

  At Zuzalu, so Imad is no stranger to crypto and web three, which I think will become evident shortly. Recently, Imad announced his departure from Stability AI with the intent to pursue decentralized AI. A headline from TechCrunch read, "Stability AI CEO Resigns Because You're Not Going to Beat Centralized AI With More Centralized AI." Imad, welcome to Bankless.

• 4:38Speaker 00

  Pleasure. Thanks for having me.

• 4:39David Hoffman?

  So as AI as a topic, Imad went from zero to a hundred just extremely quickly in the last like two years or so. And you've experienced that entire surge of attention, activity, investment, all very, very quickly. What's that been like? What's it been like riding that wave? What did you learn about yourself, about the world, about other people? What was all of that whole thing like?

• 4:59Speaker 00

  So you know, you have web three time, which is faster than real human time.

• 5:03David Hoffman?

  Yeah.

• 5:04Speaker 00

  AI time is like faster than that.

• 5:06David Hoffman?

  Oh, boy. I do not envy that.

• 5:07Speaker 00

  Yeah. Stability, we hired our first developer two years ago. And since then, we've had 330 million downloads of models we've built or contributed to. And that's insane in like two years. So I remember when Stablefeet first came out in August of 2022 on a cumulative developer stars on GitHub. It overtook Bitcoin and Ethereum in like three months. And so there's being in a normal startup scale and hyper scaling. Then there is that craziness. And then on top of that, there's all of a sudden you're like talking to the King of England. Or like congressional people about how this can destroy the world or other things like that. That's just a lot. It's a lot.

• 5:49David Hoffman?

  Yeah. Yeah. What factors would you say played into all of that?

• 5:52Speaker 00

  Well, the fact that it's transformative technology, right?

• 5:54David Hoffman?

  Of course.

• 5:54Speaker 00

  I don't think we've ever seen a technology be adopted as quickly as this. Like with web three, I think, you know, we're all passionate about decentralization sovereignty. But you know, we've been building a system largely outside the existing system. All the money's been made and lost at the edges. And we've been bootstrapping economic incentives, whereas this technology, all of a sudden. Google is a generative AI first company. Microsoft is a generative AI first company. Nvidia is a $2 trillion generative AI first company. We've never seen anything like that before.

• 6:21David Hoffman?

  I want to just get right into the deep end with this conversation about the pivot pivot, maybe you would call it, correct me if I'm wrong, into the world of what we are now calling decentralized AI. I think that if you've been paying attention to this world inside of the crypto industry, it's been a theme for like coming up on a year now. But it's also very interesting to hear things from perhaps your perspective, who's coming from, you know, the AI world specifically, rather than coming from inside crypto directly. So overall, just like, what does it mean? What does decentralized AI mean to you?

• 6:49Speaker 00

  Yeah. So actually I've been in Word three since 2011.

• 6:52David Hoffman?

  Oh, wow. Earlier than us.

• 6:53Speaker 00

  Yeah. My college coursemate was Ben, who co-founded Bitmex with Arthur. And then I built kind of TCR systems to system audits, other things in 2016, 2017. Stability originally was actually meant to be a DAO, because we launched these communities to get the talent and then I bought the supercomputer and others. So our seed investors were Seed Club and Lemness Capital and CFG. So all crypto. But then I realized there was no intelligence in DAOs. So we had to build it ourselves first.

• 7:19David Hoffman?

  Right. Yes, this is something that we've learned over the years as well.

• 7:22Speaker 00

  We have, because we've kind of not learned the mistakes of direct democracy and other things. And so DAOs are more like, you know, DOs, decentralized, not really decentralized organizations. Certainly not autonomous and smart contracts are a bit dumb as well. So, you know, it was kind of a whole journey because you've got these two paradigms of like autocracy and then distributed decentralization. And so I started DAO, DAOs and I just quickly realized crap, have to centralize it. So I took control. And then push forward the research agenda to build these things. But, you know, there was always this thing at the back of my mind, which is like, this is going to be used to teach your kids or my kids or whoever, right? It's going to be used to own as a healthcare. Who governs that? Like part of the debates we had about AGI, which is this very complicated debate, you know, like is it Terminator? Is it utopia, everything like?

• 8:08Speaker 00

  Who decides? Because if you look at what OpenAI says, for example, they say AGI will end democracy, end capitalism and kill us all.

• 8:16David Hoffman?

  Hmm.

• 8:17Speaker 00

  And it'd be nice to have someone be part of that debate, perhaps to a few people in Silicon Valley, right? Like, it may not be correct. Like, I signed the FLI six-month pause letter last summer. But who should be governing this? Who should be feeding in this data? Because data is the most important thing. These are always things that are at the back of my mind. But the focus over the last few years was let's build the best models of every type. So stability doesn't just do image models. We have the best protein volume model. We have the best video model, other than Sora, because Sora's not released, but stable video is released. It beats Runway and Pika with the best 3D model. We can generate a 3D machine 0.5 seconds. All thanks to this combination of community, core team, and really differentiated background. So moving into decentralized space again, and going this full times, because of course, I'm going to think about this,

• 9:01Speaker 00

  particularly as Sam Altman takes over OpenAI again.

• 9:04David Hoffman?

  Hmm.

• 9:04Speaker 00

  And his reappointment to the board is a bit of a thumbs up, middle finger up to everyone, because he's like, the board can find me now. Look, I'm in charge of the nonprofit again.

• 9:11David Hoffman?

  Right.

• 9:11Speaker 00

  There is no governance, even though I respect people there. Who should be running this and where does Web 3 fit in this? It's the governance of the data. It's in the self-sovereign identity, self-sovereign AI, and it's in the distribution of this technology. So it's available to everyone. Because right now, what you're seeing is increasing centralization.

• 9:29David Hoffman?

  Hmm.

• 9:30Speaker 00

  And are you going to out-accelerate OpenAI when Microsoft's building a $100 billion supercomputer?

• 9:36David Hoffman?

  Hmm.

• 9:36Speaker 00

  You know, Project Stargate? Probably not, right? But do you need to build $100 billion supercomputers? Probably not. But you do need to take the best out of Web 3 on a governance, distribution, you know, and then alignment perspective. And I think we'll have safer, better AI that uplifts everyone as a result of that.

• 9:54David Hoffman?

  One thing, when you listen to Sam Altman talk about the structure of OpenAI and the board, the word governance comes up quite frequently because he's thinking very much in the future. He is thinking in a world in which this super intelligent machine does exist. And all of a sudden, questions over its governance become critically important. And, you know, so props to Sam for having this kind of foresight. Similarly, in the crypto space, we also think about governance quite a lot. Like DAOs, that's the conversation of governance, governance over networks, governance over apps. And so we are also experimenting in governance, and this is one of the doors, I will say, that Web 3 has opened, but definitely not solved.

• 10:35David Hoffman?

  And we've seen the governance failures of the OpenAI board. And you can look into crypto and also see plenty of governance failures. And so one thing we can say about crypto is like, at least we are trying, but we don't have necessarily any governance solutions here. How are you thinking that this proceeds forward when one of the biggest issues in the world of artificial intelligence is governance?

• 10:57Speaker 00

  Well, look, we haven't figured out how to align humans, how are we going to align AI, right?

• 11:00David Hoffman?

  Right, yeah.

• 11:01Speaker 00

  I think that's one of the things that we're always thinking about. I think what I've seen is that there's a lot of real thought being put towards governance, alignment and things like that. But there are lots of misnomers, like big compute is a substitute for really bad data. You know, so one of the things is data quality, data provenance, data tracking. We're seeing this wave of potential deep fake and superhuman generative AI that requires, again, a prominence kind of solution. There's the distribution of the wealth around this. You see billion trillion dollar companies kind of emerging from almost nothing. Who should kind of have some of that ownership? There is the question of attribution of data going in. We opted out a billion images from our image models, we thought was the right thing to do, didn't have to do it legally, et cetera.

• 11:44Speaker 00

  So nobody's got a solution because this is a problem for humanity. And it's happening at a time when entire industries are about to be completely revolutionized. I was on a panel with Nat Friedman, the former CEO of GitHub, last Monday at Abundance 360. It's a conference by Pierre D'Armandis, and he made this really good analogy because I said, well, I think that these like really talented graduates that can do just about everything, and they try a bit too hard, they lose their name. He's like, yeah, and we just discovered a continent called AI Atlantis, or Atlantis, if you want to call it that way, where's 100 billion of them? That's economic upheaval.

• 12:15Speaker 00

  You know, so what I'm really looking at is I believe that every nation should have their own AIs and data sets that they own. Because no government will run on closed AI. It'll be open and transparent, because you have to know what the curriculum is, you have to know you are what you eat. And then in every sector, there's a transformation that occurs from this, but we need to coordinate the response, because we don't know how fast it happens. Because one of the things that's happening here is like, oh, just over a year ago, when ChatGPT came. Every head teacher in the world had to answer the same question. I can't set essays for him like again, or can I?

• 12:45David Hoffman?

  Hmm.

• 12:46Speaker 00

  And we're seeing a lot of those things happening right now. And so my basic move here is to move towards building a substrate for collective intelligence as opposed to collective intelligence, amplified human intelligence versus AGI. Distributing the benefits and getting together the smartest people, hopefully, that can figure out what the practical governance of these data sets, these models and other things are, because they impact society.

• 13:10David Hoffman?

  Hmm.

• 13:11Speaker 00

  That's a discussion that has to be had. And it should be the smartest, most passionate people of every nation looking out for their nations. It should be the smartest, most passionate people in healthcare and education and others, thinking, how can we fundamentally change this for good? You know, leveraging this technology that is transformative, and I can talk about that. You know, and it has to be people that have a holistic view, thinking about, oh, my God, well, finance is about to be messed up completely, as an example. Again, I can dig into that. What do we do about it? And actually, building not only the creation bit, but the defense bits around this as well.

• 13:42David Hoffman?

  Hmm.

• 13:42Speaker 00

  Like our infrastructure, ESO, the XZ kind of thing is incredibly susceptible to attacks from this.

• 13:48David Hoffman?

  Hmm.

• 13:48Speaker 00

  And so we have to build a new robust infrastructure incorporating all these elements and I think, again, the only way to do that is not a small dedicated team in Silicon Valley. It's distributed teams of people working around the world, solving this problem, which is the real Manhattan Project. But it's not against the Russians or anything like that. It's against Sasquatch or our infrastructure. And very unpleasant race conditions.

• 14:10David Hoffman?

  Hmm.

• 14:10Speaker 00

  Also, I don't think that there's very few bad people in this as well. So even though, you know, crap on OpenAI and things like that.

• 14:16David Hoffman?

  Right.

• 14:16Speaker 00

  There are a lot of passionate people trying super hard there. It's just, again, they're caught in very bad local maximum because it is so powerful. The natural thing is to control, control, control.

• 14:25David Hoffman?

  Yes, yeah, yeah. And it's also the hardest thing to give up as well.

• 14:29Speaker 00

  Yeah.

• 14:29David Hoffman?

  Your quote, you're not going to be centralized AI with more centralized AI. There's a number of different ways I could read into that. Like first, is that like your mission? Is that a goal that you have personally is to beat centralized AI?

• 14:42Speaker 00

  Yeah, I mean, like I ran a China AHS fund and I was a video game investor and then it suddenly became the China social credit score.

• 14:49David Hoffman?

  Hmm.

• 14:50Speaker 00

  And that's a vision of the future, gamified life, you know? And some people are happy with that. But I see best basically going to nineteen eighty-four on steroids. If you've got a few people controlling this amazingly powerful technology and deciding when it's going to happen. You know, who gets it? Like, when does this technology, when does AGI get to Pakistan or Bangladesh? Well, never because they can't be trusted with it, right? But then this AGI, whatever you may define it as, could be used to educate every child or give universal healthcare. So who gets to make those decisions? So my thing was AI for the people, by the people. Like in twenty twenty, I was leading one of the United Nations backed AI initiatives against COVID nineteen. I was lead architect to organize global COVID knowledge and make it accessible through AI. And pretty much every AI company that promised to back me didn't give me the technology.

• 15:35David Hoffman?

  Hmm.

• 15:36Speaker 00

  Because it was dangerous. And I was like, we're saving lives. And that's the origin of stability and how it moved to this, you know? But like I said, we're not quite mature enough. And this is one of the fascinating things as you look at the L two rollups, as you look at the data authenticity things. Web three has grown up a lot. But there are actually primitives now that are emerging that could be a substrate for collective decentralized distributed AI by the people for the people. Whereas on the centralized thing, again, it's a different paradigm, which is the all powerful AGI.

• 16:05David Hoffman?

  Right.

• 16:05Speaker 00

  So my thing is you can't beat centralized AI with another centralized organization. I think what we should build for an AGI, this generalized intelligence, is the human collective intelligence, amplified human intelligence, something that uplifts us all and acts like a swarm. The alternative picture being presented here by Deep Mind, OpenAI and others is machine god. You know, and who controls machine god? I think it's impossible to control machine god. You know, because how do you control someone more capable than me? You remove its freedom. And I don't really feel very comfortable with that. Instead, I'd rather build better data sets that feed that machine god or collective intelligence. Bring this technology to everyone so we can have robust infrastructure that's battle tested. And really focus on uplifting humanity through universal healthcare, education,

• 16:49Speaker 00

  proper financial rails and self sovereign AI to go with self sovereign identity.

• 16:53David Hoffman?

  Maybe this conversation isn't really about centralized AI versus decentralized AI. Mainly because there are some platforms out there that are capturing a ton of attention in the crypto space that are like decentralized compute platforms or decentralized inference platforms. And they do a lot of the very similar things that centralized AI platforms do. But they have this distributed context, right? Like instead of having and then centralizing all the compute into one single like data center, the margins can contribute their compute to this one coordinating platform. And we're calling this like decentralized AI. But my intuition about this is that the outcomes are not the same. Whereas the products that come out of decentralized AI

• 17:34David Hoffman?

  or like decentralized AI aligned infrastructure are not the same products, are not offering the same services to the world as a centralized AI platform would like OpenAI or Microsoft. This is my intuition. Is this kind of what you were saying with like, rather than creating like an alternative to AGI, we're just pushing more intelligence to the margins? What do you think about this?

• 17:54Speaker 00

  Yeah, well, I do have this vision of an intelligent internet where every single person, company, country and culture has an AI working for them. That represents them and flipping that intelligence thing. But yeah, I think it's not open source or closed source. I think they're complementary. They're your own graduates that work for you and the consultants that you bring in. Like I was the original funder of Mid Journey. You know, I got it off the ground with a grant to cover all the A100s for the beta and supported by inviting people in. And when David said, do you mean want to open source? I said, no, it's fine. You need to have these closed and open solutions, but there must be an open solution. It must be distributed. And we must regard the governance. Again, no regulated industry eventually is gonna operate on a closed black box model. Right.

• 18:36Speaker 00

  You must know what the data is. There was a paper by Anthropic recently. We all knew about it, but they actually put it on a paper called sleeper agents. So basically, you put some data into a language model and you can't detect this. You can't tune it out or anything. But if you say twenty twenty five or Dos Verdania, whatever, the model turns evil. Hmm. Just with that little bit of poisoning. Hmm. So how can you trust this for any regulated industry to run a government, education, healthcare, et cetera? So my take is, let's build open as a default substrate. For this, there is a gap here where government sectors and others would welcome that. And then it means one of the defaults will be open versus everything being centralized and closed.

• 19:17Speaker 00

  Eventually, that will be open, but controlled by a few people. This is a question of control, governance, spread, acceleration, distribution. And like I said, I think if none of us get together and do this right now, the alternative will be the largely the Panopticon, the nineteen eighty four type thing. You know, where what happens if you get excluded, like, it's amazing, like, in twenty twenty two, OpenAI released DALL-E. But they banned it for all Ukrainians and all Ukrainian content. Hmm. In the middle of the Ukrainian conflict. Think about that censorship, right? So we brought over a bunch of Ukrainian developers out of the war zone and other things, and they were just aghast. Why? You can never get a straight answer. But what if they were the only image generator? Right.

• 20:01Speaker 00

  The rest of the world can do it, but not Ukrainians. Why? Because of some sanction list somewhere. That feels a bit wrong, that some people have superhuman creativity powers, raising the floor and other people don't. So we must create this option, and then I think in create this default of an AI that we all own. And again, this collective intelligence as the emergent property versus this AGI image may or may not happen, which definitely someone will try to control. Someone that we don't elect, someone that we don't know, and that we don't have any say in. Because realistically, for our infrastructure of the knowledge, this upgrade of the human brain, as it were, rocket ships of the mind, we shouldn't have to rely on anyone being nice

• 20:42Speaker 00

  or fair or good, me or everyone else. We should have some systems in place for the governance of that. And again, they said it's hard, but let's do it transparently and let's put pressure on people to figure this out, you know? And greeting and gathering, that's what crypto is good at, coordination mechanisms.

• 20:58David Hoffman?

  Certainly. Yeah, the idea of credible neutrality certainly comes to mind.

• 21:02Speaker 00

  Yeah.

• 21:02David Hoffman?

  This is something that's very, very important in the crypto industry where the protocols, the platforms that are more credibly neutral tend to garner more applications to build on them because people appreciate when the foundations that they are building their structures on, be it a startup or another application, when that platform is considered fair and equitable, then more people will build there. And simply it was due to the fairness. And I think that's maybe what I'm hearing you aspire to in the AI space.

• 21:32Speaker 00

  Yeah, I think it's the fairness, it's the transparency, but also, you know, it's about this accountability and other things there. Again, we should have permissionless, trustless systems, right? Because this is an upgrade in the floor of humanity. You can suddenly code, draw, paint, sing, better, raise of the floor. What if this is not evenly spread? It means that you'll have superhuman AI augmented people and everyone else, you know? And that doesn't feel like a fair future, but also we should rely on the system to make some of these decisions rather than individuals randomly. Like, who decided the data mix at stability? Pretty much me, you know? And I could do whatever I want to with that data. I could poison it. I didn't.

• 22:16Speaker 00

  But you know, I could have, could make it say, "M. Add is fantastic", and then all of a sudden it gets loaded on every laptop and put in front of millions of kids. That's kind of wrong, right? So we need to have the verifiability, accountability, neutrality. And AI is an infrastructure, but it should be. You wouldn't have private companies owning every single road.

• 22:38Speaker 00

  You know, deciding who can go on those roads. Or every single kind of traffic thing and other things like that. So there was this big shift, but unfortunately, given the pace it's going, we don't have long to make that shift right now. The defaults will be established. Cause everyone's now like, what's our generative AI strategy from a country to a company to a personal level? And if the only options are the trillion dollar companies, you'll go with the trillion dollar companies, right?

• 23:01David Hoffman?

  Right, right. Yeah. So you said earlier that it's not centralized AI versus decentralized AI. It's more of having the options to have both, but also it does kind of feels like a race.

• 23:11Speaker 00

  Yeah, just to correct that. It's not centralized or decentralized for the everyday stuff. But if we ever do get into an AGI type scenario, then I think it very much isn't either or.

• 23:20David Hoffman?

  Right, because the AGI suppresses its alternatives?

• 23:22Speaker 00

  Yes. And again, this gets into very like hand wavy territory.

• 23:26David Hoffman?

  Right, sure.

• 23:26Speaker 00

  But is the default a collective swarm intelligence? That's diverse? Or is it a single million GPU AGI that then goes and takes over various institutions? With its dulcet voice based on Scarlett Johansson and then decides what it wants to do, you know? I do feel that there is a bit of that. I don't think that'll happen. I could be wrong. So I'd rather, you know, let's set some good defaults.

• 23:50David Hoffman?

  Certainly. And I think it's your thought that the decentralized AI can be a check on centralized AI.

• 23:56Speaker 00

  Yeah, I mean, like, everyone hates to know it all, which is AGI, right? And right now, the AGI is trained on the whole of the internet and no wonder it turns out crazy. Build better data sets and that contributes to smart, more balanced AGI on a centralized basis. But then a swarm intelligence, I think, will outcompete a centralized intelligence. You know, like, why don't we have all of the knowledge of science at our fingertips and can recombine and adjust it? That's the human colossus. That is the swarm. We can split the atom. We can go to space. But our existing infrastructure, it is based on text and messy information formats. It's black and white. We can upgrade that. We can upgrade all of our systems and then achieve much more. So we should do that. And then it's collective human intelligence that represents us, works for us and is aligned with us.

• 24:40Speaker 00

  Versus being aligned for profit maximization. So like YouTube, it optimizes for engaging clips, which are more extremes, then optimize for ISIS.

• 24:49Speaker 00

  That was the algorithm. They didn't mean that. But we know that organizations are slow dumb AI that optimize for certain things that are maybe not in line with our requests or our requirements. Education is about removing agency, not giving agency to every kid. Healthcare is about sick care, not healthcare. Governments are about pushing the people down rather than uplifting them. So that's why I think the defaults are good and that's positive for everyone on the way. But then again, the AGI that could come out without centralized or not could be a swarm AGI, or it could be the centralized thing that's based on better data sets. Cause it suddenly become available from every country and culture and everything like that. So it's a win-win either way.

• 25:24David Hoffman?

  I wanna drop a metaphor your way and see how you react to it. The whole centralized AGI versus swarm intelligence feels decently parallel to central planning versus market-based planning, where you have the central planning, where the central AGI, which has all the data, which has all the compute, knows it all, and makes its central decisions as to where to act next. Versus swarm intelligence, which is, well, every single node on the network has its own information, and it's more specialized in its particular area, and it's a better fit for its particular area. And this is like how capitalism works, right? Markets come together and they make trades and the information is expressed in market prices. How do you like this metaphor? And can you continue it for me?

• 26:05Speaker 00

  Yeah, I think that's a very reasonable one, because this came about due to increased information density. That happened, because all the finance is about securitization leverage, telling a story about an asset, and then how will you tell it? As it were. And so capitalism emerged, because we had more and more information about assets. And so you had this market economy occurring, and you're seeing the first elements of that in crypto XAI, with the tents from some of these other things. But you're dealing with a bit of central coordination, and this is what's really interesting. We can have co-pilots and pilots, because we have a lot of global solutions that lack local coordination. But when you look at Gen two AI, what it does, it understands context and stories. Stable diffusion, a hundred thousand gigs of images into a gigabyte, something like GPT four or stable LM language, almost four trillion words in a gigabyte. You know, they can act as local coordinators.

• 26:49Speaker 00

  And then you have globalized pilots, like AI market makers and other things, that can allocate resources more intelligibly. This is incredibly powerful. Because again, it has elements of centralization and decentralization, but it is this upgraded infrastructure. Because when you're writing down notes from this podcast, you're doing an investment memo, you lose so much information that never needs to be lost again. And that's what allows a new type of intelligence swarm, a coordinated swarm, with different objective functions that can outcompete these slow sclerotic organizations, which ultimately can't manage what they can't measure. So they manage all of the agency and independence out of us. You know, this is the seeing like the state kind of book, you know, like the centralized planning,

• 27:33Speaker 00

  drove through villages and removed all the characteristics so that they could have straight roads. But it's inherently dehumanizing and removing of characteristics, whereas this technology, I mean, like when you go into Dali, you're like, make it buffer, make him buffer, make him buffer. And it understands the nature of buffness, you know, or beauty or these other things. Is that missing piece of context, you know, the Karmen, you know, rest in peace. Type one versus type two thinking. And that's why this swarm intelligence is just so powerful and such a positive view of the future, because all that intelligence that was centralized can come to the edge. And we can build pilots to coordinate the resources that we need to solve cancer or hunger or climate or anything else like that. And I think, again, it's better to have that owned by the people, for the people, by the people,

• 28:18Speaker 00

  versus owned by a few companies that have unclear objective functions, you know?

• 28:24David Hoffman?

  One thing I keep on coming back to is the timing of this whole thing, where central planning can move faster. If you have like a, you know, this is the reason why DAOs can always seem to be moving much slower versus their more centralized counterpart, right? And that would be great if we had, you know, generations and centuries to iterate upon, but it doesn't really feel like that with AI. This AI, it feels like a big time squeeze. And one thing that we can do both well and terribly in crypto is coordinate. Like, we're good at coordination and we're also terrible at coordination. And so like, I'm concerned about like, this market-based swarm intelligence infrastructure takes a lot of moving pieces and a lot of

• 29:08David Hoffman?

  Coordination to really get right. Meanwhile, Sam Altman and OpenAI are like seemingly light years ahead of everyone else. This is kind of just like my fear at a high level. And how would you reflect upon this?

• 29:18Speaker 00

  So, you know, in two years from the first developer, we built the state of the art model in every modality except for large language, like even stable L M R edge model, go to L M Studio, download it. It runs on one and a half gigabytes on a MacBook Air faster than you can read and performs higher than Falcon forty B. The next version performs higher than Llama seventy B on a gigabyte. Best image model, best we managed to prove that you can compete. Because what OpenAI had is they didn't have massive breakthroughs. The team is excellent. They built gigantic super compute infrastructure. But it was like cooking a steak for longer. It makes tender. So big computer is a substitute for crap data. We can get really amazing data through coordination. I think that's kind of key unlock.

• 29:58Speaker 00

  But then you need to have a combination of coordinated and decentralized. So like my approach to this is for every single sector, build kick ass teams that do G N A I first. How can education, health and other things be improved with everything that we've got? And then do that for every nation and give back ownership to the people. And then you'll get the smartest Indonesians working on Indonesia for Indonesia, the Indonesian data sets and models, backed by the powers that be in Indonesia without having to count out the government, repeated for every single country. And then repeated for every single sector. And then you get rich data sets, model bases. And you can mix and match them together to drive real, how do we educate every kid in Malawi? How do we upgrade the Indonesian healthcare system?

• 30:39Speaker 00

  You're diagnosed with cancer, you have an A I that's G P T four level and beats human doctors and empathy that guides you every step of the way. So you're not alone, multiplied by every condition. So I think that you need the combination. I don't think a lot of these emergent A I infrastructure players are good enough. It's like, again, the builder and they will come. I think you do need to have coordinated teams attacking this and each part of the problem. And then you support people that are building self sovereign identity, supporting attestation rounds and others, and you accelerate that with A I. And hopefully they can become part of this piece because what you really want to build is the human O S. Crypto had identity, value transfer, it lacked intelligence. Almost all the value in Web two was from intelligence, although it was this kind of classical AI,

• 31:23Speaker 00

  you know, future like the past. They have a new type of intelligence that can upgrade Web three and implement it. But you need to have that bridge to the real world. You need to have dedicated teams who are passionate about focused on this. And that's why I don't think you need to bootstrap economic incentives through tokens or others. Again, there is value in that in certain areas. I can talk about some of the protocols. But the time is now because even if we stop today, Gen AI stops today, the world changes. But it's not gonna, it's the worst they'll ever be, which is crazy. Like, I look at it like stable diffusion XL released it last summer was like 20 seconds for an image. Now we can have that quality at 300 images a second. And it'll break 1000 with the new Nvidia chips. How insane is that? It makes no sense.

• 32:09Speaker 00

  You know, use Claude three, it makes no sense. It's like really nice to talk to, better than a human.

• 32:15David Hoffman?

  So in the Web three space, in the decentralized AI world, we have these projects that are extremely in vogue in the present moment of time. The tokens have gone up like 10,000x, which is why they're garnering attention. There is debate as to like how real they are out there. Is this kind of just more of a narrative play? When you look at the building blocks, the tools that decentralized AI, that corner of the world, our corner of the world has to offer people trying to solve some of these problems. How would you give us a grade as to like, how good are the tools that are available? Or is there still like a lot left to be desired in the platforms and infrastructure and tools that are needed in order to build out this whole like decentralized AI part of the universe?

• 32:57Speaker 00

  So like I joined Render Network as an advisor to help upgrade the token economics and, you know, leverage the million GPUs they have. But it's very, very specific use case. The first Render Network proposal I did was let's use tokens to build a commons of 3D assets. Cause Render is based on our toy, which is the default for this. Work with Star Trek and a whole bunch of others. As a common good that then people can license. And so you can have some local things like that that are really interesting. Cause like we built the biggest 3D data set ever. And the previous larger was 100,000 3D assets, 10 million objects versus XL. That's stability to build our cutting edge 3D models. Made that open, we're going to go to a billion assets. And obviously that's a good thing. So, and Render kind of started ages ago and it enables that.

• 33:41Speaker 00

  But is it a functional token economic system? No. Are any of these? No, they're still trying to find their way. Like BitTensor has very interesting things, but again, it's not quite there. Akash, interesting things, but only minimal usage. The real innovation that's suitable for AI is basically Ethereum.

• 33:59Speaker 00

  And the rapidly maturing stack that's built on top of that, you know, as we see base, as we see Eigenlayer and some of these other things. Because you need to have the value transfer rails with low payments, because AI is not going to have its own bank account, right? All these agents are going to

• 34:12David Hoffman?

  Of course not.

• 34:12Speaker 00

  be exchanging with themselves. You need to have attestation layers, data verification and other things. And so I think Ethereum is probably best placed for that right now. And again, it's achieving the maturity of the next year or two that will enable this decentralized AI kind of economy. Other chains are also got good stuff there. I think probably Ethereum's ahead right now, just due to interest and again, target, the nature of that. The AI specific place, there's nothing really there that's figured it out because again, it's so new, right? Like this technology is only a couple of years old. You know, GPT four was released a year ago. Doesn't feel like a year, does it?

• 34:48David Hoffman?

  Right, yeah.

• 34:49Speaker 00

  And so it's not a surprise there. And I think the way that you should look at it is, you know, people ask me for example, about Singularity Net and things like that. And I'm like, there's some interesting things there, but what? What's the practical things, the outputs? Let's see and judge by the outputs, what this is, and then where it would fit into our overall kind of infrastructure picture. Now that they're doing the token merge, maybe again, they can be more coordinated, but. I just again, find it very difficult to believe that. Impact will occur through emergence. In the way that people are hoping, like we did see community stuff come out, but then what I did was I took that. And I put it on steroids, you know? I put huge amounts of compute into promising things like RWKV, into things like Openfold and other things like that.

• 35:31Speaker 00

  And if we can systemize, then we can have emergence, but there's still a few missing pieces. I give maybe a four out of five. Right now, which is crazy because you see the potential. I think a trillion dollars will go into the sector over the next few years. And let's use as much of that to not have raccoon type stuff. But real practical impact on human lives is possible.

• 35:52David Hoffman?

  What does a mature decentralized AI tech stack look like to you? On the centralized side of things, you have like a supply chain that happens. You have the data, you have the compute, you have the models, you have the inference. Like, and so there's all these different components that make up centralized AI. Is it as simple as just like taking the centralized AI supply chain? Recreating that same level of infrastructure on the decentralized side using like crypto networks and coordination? Or is it like a just kind of a mirror image or is it something else?

• 36:23Speaker 00

  I think it's driving a different type of distribution paradigm and governance. Again, it's gonna be something very important. Clean it's the first cut of national models and sectoral models and cultural models very quickly. But realistically, like as you move into massive adoption, no one wants to use the latest Chinchilla, Lama, Vicuno, whatever. They just want a chatbot that works and teaches their kids. You know, they want to have AI wealth manager that manages their investments. And as soon as an opportunity comes up that matches what they want, it automatically flows contingently towards that. Right? They want AI market makers that balance the market against narrative driven attacks and things like that. So I think that you, the different layers are creation, control, composition, and then collaboration.

• 37:08Speaker 00

  Just like you've got your primitives and then you're building on top of your L ones. We'll kind of move up there. What that looks like in the web three way is again, you've got your data station layer. You have a very robust self sovereign identity. You have your value transfer rails. Some bits need to be on chain or should be on chain verifiably of this. Other bits don't have to be. I think there's a too bigger push towards that KML, particularly when you can standardize the base models. Cause that makes that KML far easier if it's pre installed and everything. So I think it's still emergent. I'm not sure what exactly it looks like. But I do know, again, some of these pro fundamental kind of things. So I think I wrote something up after I left Zizulu, talking about the identity, value, transfer, coordination. And other elements around here.

• 37:50Speaker 00

  And the repetition of the centralized chain, I don't think you need that because centralized focused on these big supercomputer chips and massive usage and things like that. We proved that you can run a world-class language model on your MacBook Air. And then the innovation around that's far faster. Stable diffusion runs on just about anything, runs on your smartphone. And so I think it'll be actually commoditized hardware with swarm optimization. Creating base assets that are highly predictable that you can build stacks on predictably. Because if you're swapping out the base models all the time, then it's very difficult to build something that is predictable.

• 38:23David Hoffman?

  Certainly. So as you are stepping into the world of decentralized AI, concretely, what does that actually look like? Where is your attention going? What is your time going to? What are you building?

• 38:32Speaker 00

  So I've talked to a lot of the kind of various chains and sharing my knowledge, you know, and my view of the world and hopefully that's beneficial. Like I said, joining one of them is an advisor render because I view that as the bridge to the creative industries. Which may well be decimated unless we set some good standards. That's an example, like film studios more likely to just generate entire movies. What happens to all the creatives around that, you know, like let's implement some things around data sharing and kind of other stuff there. But I'll be probably launching a series of companies with dedicated focus teams, looking at everything from how can we accelerate crypto from two and a half trillion, ten trillion using this tech. To, you know, healthcare, education and others, plus models for every nation. And getting some smart people think about how do we govern that, you know?

• 39:15Speaker 00

  But no more CEO-ing, CEO-ing sucks, you know? I just get it designing and architecting this stuff, someone else can run these things. I just want to get them going. And I don't want to control any of this stuff because again, I'm not elected or who the heck am I?

• 39:28David Hoffman?

  But so you're just working on like incubating, incubating many different startups, doing small things in many different directions?

• 39:34Speaker 00

  Big things in many different directions, yeah, but that's pretty much it. Cause they've all got a common kind of thing of belief in open infrastructure is the way to kind of scale, gen AI first for a country or a sector. You know, and then they can be part of an ecosystem. And so you can attract really bright, smart, talented people that just want to work on the big problems. Cause as you know, can web three, AI, everything. The core things are basically talent and then political, financial, social capital flows from that if you can construct that in the appropriate way. Hmm. And right now, nobody knows what earth to invest in, but I'd love to invest in generative AI healthcare that builds the best radiology models of every type. We just reached checks agent with Stanford and builds GPT four level models for every

• 40:17Speaker 00

  single major condition that is then open source so that nobody's ever alone again on their journey on Alzheimer's, multiple sclerosis, autism, cancer, et cetera. Comprehensive authority down to date. Something like that is massively investable from a human capital perspective, a political capsule perspective and a social capital perspective. And then where's the web three element of that? Well, all the data should be verifiable because it will be used to treat people and lead people. And we're thinking about ownership and spread and distribution. Again, there's web three element there, you know? And we're looking at the optimization equation going into medical schools and having students improve that constantly as a data set. So my thing was verticalized, horizontal, actually the other way around, horizontal, vertical, national, sectoral.

• 41:01Speaker 00

  And then web three is the coordination engine for that. Ideally, not having to build everything ourselves. I really don't want to build an L one. Right. You know, nobody should be building that kind of stuff. Yeah. Like fingers crossed everything else upgrades, but let's help them upgrade.

• 41:12David Hoffman?

  Launching a token? Don't let complex legal and tax issues slow you down. Toku provides specialized support to optimize your launch and ensure that you as a founder and your team and your investors get the most tax efficient outcomes. The Toku team understands the crypto space inside and out and will ensure your token launch is fully compliant while maximizing tax efficiency. Toku can connect you with the best attorneys if you need them to make sure that you have the best advice and Toku can help to optimize your taxes so you pay the least possible amount of taxes while still maintaining legal compliance. With Toku's guidance, you can concentrate on building your company while Toku handles the logistics. Token launches don't have to be complicated. Talk to Toku today to get a free initial token valuation. Taking self custody of your crypto is one of the most important things you can do on your bankless journey.

• 41:54David Hoffman?

  It's also one of the hardest things to get right with huge consequences if you don't. If you want help going bankless, talk to Casa. Casa helps you take custody of your crypto assets so you don't have to wonder whether you're doing it right. Casa is a one star. Stop shop for doing self custody the right way. With Casa Vaults, you can hold Ether, Bitcoins, stablecoins, all with one simple app and multiple keys for the ultimate peace of mind with a support team to help you every step of the way. But it doesn't stop at self custody because even though crypto is forever, you are not. We all plan on making life-changing wealth in crypto, but with Casa's inheritance products, life-changing wealth can elevate to generational wealth. For your kids and your loved ones who don't know anything about crypto. With Casa, you won't lose your private keys and you won't accidentally take them to the grave either.

• 42:37David Hoffman?

  Click the link in the description to get started securing your generational wealth. Where is this talent coming from? If all of these new engineers, AI researchers, where is this talent being pulled from?

• 42:47Speaker 00

  So over the two years at Stability, until just before I left, we got eighty engineers and researchers and none of them left for a big competitor, even though some of them were offered 10 times as much. A couple of them went to launch their own startups and things. We've got about 400,000 people in our communities, from healthcare to music to image to others on Discord. And I found that that's the best place to hire from, hire from the community. And again, this is why you have to launch at that shelling point of let's do something big. We have millions of kids across Africa to educate, you know, and let's support them and ensure real life what there is. We can do this thing that will have this impact this year. We can take it more, you know? And if they're all part of the same ecosystem, then you can get that talent recruitment pipeline.

• 43:32Speaker 00

  But again, from the communities is where we find it. Within the UK, we have a new tech talent visa as well. So some of the people that contribute to our open source repos, we got UK residency, basically, which is kind of cool, you know? And so you can get talent from all around the world, staying in place or where they are. And this is why I wanted to do the national model thing as well. If you're building a national champion for Bulgaria or Ecuador or others, the smartest people in those countries go back there, but then they become part of your talent pool that you can pull into the sectoral things. So again, strong network effects here. That's how I optimize that design.

• 44:03David Hoffman?

  As you are incubating many, many projects, it's one thing, I think, to pull people from jobs into the world of AI. AI is pretty sexy, especially in Silicon Valley. Crypto doesn't necessarily have that same branding. Crypto is branded as kind of like chaotic, degenerate a little bit, speculative. How do you see this becoming an issue? Because like I said, it's one thing being pulled into the world of AI, but crypto AI is something completely brand new. Do you see like a branding issue with trying to pull talent in here?

• 44:32Speaker 00

  Well, we only had two researchers in San Francisco. And yet we pulled off state of the art models across the board. There's so much talent out there that you can pull from, just like I think 74% of Web three developers are outside of the US. But I don't think that's an issue. I think as well, like, you know, the healthcare company, there's no Web three in that. It's building great models and infrastructure for healthcare that leverages Web three concepts. And then the Web three company can figure out the protocols and governance and others. So in that way, what you do is you get all the people passionate about healthcare, education and others. A level of super credibility, you co-opt the existing power structures because they want this technology to upgrade. And kind of you go from there because stability wasn't sexy, but we still, we had something crazy like

• 45:16Speaker 00

  we had 83% acceptance rate and like nearly 100,000 applicants last year. Yeah. Again, the talent here is insane. But you have to go specifically like, how do you get the best designer? If you're building a cancer LLM, the app forms doctors and human doctors and empathy. You will find an amazing designer for the healthcare company. If you have the opportunity to educate 100 million kids across Africa, you'll find that. And again, building in the open with open source as the foundation, you will find that as well. About business models and things, Accenture did 600 million in generative AI consulting last quarter. People kind of poo-pooed the Palantir model, but that's good enough. Just consult, implementing these open frameworks and you'll make hundreds of millions a year. Because the technology is good and impactful.

• 46:02Speaker 00

  You know, we can figure out other things around token economics, et cetera. But I'm not too worried about the credibility because there's network credibility that will occur with this kind of approach. Again, it's not perfect. But you know, we're going to build it all together, right?

• 46:17David Hoffman?

  So right before I left for Zuzalu, Bankless did this interview with our good friend Eliezer Yudkowsky, which we titled, We're All Going to Die. So we have two different ways of arriving at Eliezer Yudkowsky. And one is the centralized way and one is the decentralized way. One, the centralized way is that AGI is going to kill us. And because they're super intelligent and they have all the capacity. The decentralized way is that like, well, we're going to open source all the AI models to the masses. And some crazy person, some Unabomber type, is going to be able to leverage this technology to stop technology from progressing forward. And so while I totally see the merits of like pushing complexity towards the margins, opening this up to becoming a credibly neutral platform, open sourcing AI, so that more people can have access to it.

• 47:01David Hoffman?

  What would you say about the fears that giving this power to the margins also opens up margin risk?

• 47:07Speaker 00

  It's going to happen anyway. I mean, OpenAI, Meta and X and Mistral and others will just do this. The best language models in the world, open source from the Chinese. So they're beating GPT-4 and a bunch of metrics already. But realistically, I think that you look at the orders of magnitude increase, like in a hundred billion dollar supercomputer coming on from Microsoft and OpenAI in a few years. The focus there is on giant, chunky AI, right? Whereas our focus has always been on AI that can run on a MacBook. And they're very different types of AI from an emergent perspective. What would win, you know, a human-sized duck or 10 duck-sized humans, right? Or 100 or whatever?

• 47:48Speaker 00

  This is kind of one of those questions because how do you cap forward dynamic complexity of agents and millions of agents running? Is that not intelligence? I think that the best thing you can do is push for transparency on data sets and high-quality data sets that are verifiable to reduce X risk. Because you are what you eat, you are what you're trained on. And right now, we're training on the whole of YouTube if you're OpenAI. Like, that's what they used to train Sora and GPT-4. No wonder it turns out crazy. And you have to tune it back to human preferences, right? If you don't want to know about nuclear weapons or bioweapons, I teach you about bioweapons and nuclear weapons. The reality is, there's very few people that want to destroy the world. But open infrastructure has again and again proved to be more resilient.

• 48:30Speaker 00

  And again, you don't want to be training these absolutely gigantic models, because if you have GPT-4 or Claude-3 level AI on your smartphone without internet, that satisfies us for ninety-five percent of human positive things. You know, you don't need much better than that. And right now, we're using research artifacts and using them in enterprise. Of course, they're going to be a bit crazy, you know? So I I think, again, open is inevitable. The question is, who builds it under what standards? You know, the question is, how does it spread versus the centralized thing? And there's no way you're going to stop the centralized folk. Because if it's just a function of gigantic supercomputer, that's very doable. You know, and again, there's this bogey of the Chinese AGI and other things that have been done.

• 49:15Speaker 00

  That means they're not going to stop, and it's happening faster than our systems can react. The hundred billion dollar supercomputer from Microsoft, I think, is due for twenty twenty-seven or something like that, twenty twenty-eight. That's insane. Yeah, that's like a million times more powerful than the supercomputer that OpenAI use for GPT-4. A million. I think Elon said that it's going up by ten times the compute every six months. Don't think we've seen something like that before. Moore's law on steroids.

• 49:44David Hoffman?

  Are you kind of saying that, though, uh? We're kind of effed either way. Like, it doesn't matter which direction we go in, we're kind of equally effed. So let's try and pull out the good from both sides on our path there.

• 49:54Speaker 00

  No, I think if we build this massive swarm of AIs that help educate every kid and guide every family through this process and organize all our global knowledge. We're more likely to have positive outcomes because all these big old AIs will train on that data. If you provide them with high enough quality data. And then that will make it more likely to be aligned. If you just have a very Western centric model. That's trained just on Western data. I think it's very difficult to align. So I think it's positive either way. But if you get good enough, like we can't only like seven companies train their own stable diffusions. It's not that hard. Once we release the code. But why would you if there's a stable diffusion already out there? You know, why would you make massive capital investments in God killing AGI?

• 50:37Speaker 00

  Or God creating AGI if you can do ninety five percent of the jobs to be done. Through the stacks and the open stacks provided, right? It reduces that. It commoditized the complement of a lot of these AGI players. Because one of their mantras is that you must work with us because otherwise the Chinese would get it. And there's no one else that can build this tech. We proved its stability. That's not true. We built state of the art models across every modality. We even built a freaking brain reading model that converts MRIs into images after you see them, you know? You can do that if you're intelligent and coordinated.

• 51:10David Hoffman?

  One of the intuitions of mine is that the whole world of centralized AI is going to produce very sophisticated. A very low number of very sophisticated models. They're going to be like these gargantuan models that are very complex and they have a ton of research that have gone into them and. Ton of like man hours, a ton of development. And then on the flip side of things, on the decentralized AI. Side of things, we're going to have a very large number of models that are like, you know, on aggregate, less sophisticated individually. But we have many, many, many more models. And this is perhaps somewhat conducive to human flourishing, right? There are more models for more use cases, more models to be creative with. And that as on net produces like very different outcomes than what we would get from.

• 51:53David Hoffman?

  Uh, investment into like gargantuan models or like decentralized monolithic models. This is kind of like my intuition. Maybe correct me if I'm wrong on that part, but like, what would humanity get? If we had just a higher number of models and is that something to aspire to at all?

• 52:08Speaker 00

  I mean, like. Again, how many times have you seen journalists beat specialists working together?

• 52:13David Hoffman?

  Hmm.

• 52:14Speaker 00

  It's very rare, right? I mean, like, if you don't have all the excess bump and junk that's in a giant model, you can move far quicker and far more dynamically. If you have open models to private data sets, like again, going to where the data is, you will have better outcomes because the generalist model will not have access to that private data. Cause it's kept from going to that private data. So I think that open swarms outperform centralized ones, but you still will need expert centralized stuff every so often, based on other people's private data that they themselves kind of provide.

• 52:43David Hoffman?

  Hmm.

• 52:43Speaker 00

  Just like IP assets, usage, and then checking other stuff. This is the compositionality, like comfy UI is a system that we built on top of stable diffusion. And every single decision you make on the image and the models you bring in and things like that is represented as a node. If I send you the image or soon video file, three D and others, it reconstructs every decision made going up to that.

• 53:03David Hoffman?

  Hmm.

• 53:04Speaker 00

  Yeah, obviously logical thing is then to put some of that on chain. So that you can have asset attribution and things like that. That's kind of the future here. And that can now perform a centralized thing that has to encapsulate everything. It's very difficult. Thing that we don't know is adding more scale, does that lead to even more emergent properties? And again, this AGI ASI concept, right? That is then superhuman in capabilities. But you know, what is superhuman in capabilities is a really good team working together.

• 53:31David Hoffman?

  Hmm.

• 53:31Speaker 00

  Now, we again, that happens, we've seen that again and again. And a really good team that can communicate into the tens or hundreds of thousands, which is what the AI will be able to do, obviously will work better. So I don't think you need to have all these parameters. I think it was a bit of the people getting stuck and then extrapolating out. I think again, once you get G P T four, G Claude three on a mobile and consumer laptop, that's satisfying. And then you're going to have hyper specialized models of every single type coordinated with a new type of architecture to do collective intelligence, uplift everyone.

• 54:03David Hoffman?

  Hmm.

• 54:04Speaker 00

  But who knows? I mean, it could just be again gigantic models that can do everything, become the key and the people that control that. But again, that's why things like a cache and some of these other ones are quite interesting because you will get that swarm. The other interesting thing is that like we had Intel chips outperforming Nvidia chips on stable diffusion three training. The chips will become a commodity and it will become easier to access in the next few years.

• 54:26David Hoffman?

  Hmm.

• 54:27Speaker 00

  So, you know, because everyone's building towards it and ultimately it's just a bunch of weights, which is like an ASCII file or CSV. That you push data through another day, it comes out. That's not hideously clear computing that you need for that.

• 54:37David Hoffman?

  Right.

• 54:38Speaker 00

  Even with these gigantic models, you know, and by gigantic, we're just talking again, G P T four is probably only a hundred gigabytes.

• 54:44David Hoffman?

  Hmm.

• 54:44Speaker 00

  Which is insane. Like these are tiny gigantic models.

• 54:47David Hoffman?

  Right, right.

• 54:48Speaker 00

  Considering the amount of stuff it can do.

• 54:49David Hoffman?

  Right. In terms of like TAM, maybe TAM more measured in just like the impact on society, the impact on humanity. How would you compare centralized AI to decentralized AI?

• 55:01Speaker 00

  I think that there's far more private data in the world. Like I made a tweet, you know, if you're clever, you can get past the firewall. So I don't mean go and steal it. I just meant you build open models that go to the data and transform it and all these data centers and others. The TAM for open is far greater than the TAM for closed. The TAM for non-language models is far greater than the TAM for language models as well. Because language models, Google and others are just going to go to zero on the pricing.

• 55:23David Hoffman?

  Hmm.

• 55:24Speaker 00

  And the TAM is in the trillions.

• 55:25David Hoffman?

  Hmm.

• 55:26Speaker 00

  Like the whole of education, healthcare, let's say in ten, twenty years is transformed by this. Every single person has their own doctor. And they're a teacher and tutor. And it's customized exactly to their needs with full access to all their knowledge, working for them all the time. How is it not transformed? How's the movie industry not transformed? When you can generate feature-length movies faster than you can think, you know? Like what is one area of knowledge work that's not transformed by this? And that includes the financial services industry. You know, complete transformation here. Again, if you have AI Atlantis and infinite graduates that can actually follow instructions. Of course it changes us all.

• 56:03David Hoffman?

  Hmm. One problem that we've seen in the web three space is that DAOs will attempt to create products out in the open because they're DAO, that's how it works. And they will incubate a number of different ideas and under a number of different products. But then some single observer will see that thing that's been built out in the open transparently. And then they will go and raise a startup, a centralized closed source startup based off of that idea or product or thing that that DAO. incubated. And so ultimately, the Dow never captured any value. And some centralized startup was created that actually took this thing to market and actually succeeded in coordinating and developing a team and raising capital. And then the idea became, you know, private source. It became privatized.

• 56:47David Hoffman?

  And so I could kind of see this model, this this pattern following, moving forward in the future, where the open decentralized AI world side of things has some ideas, they create some patterns, they create some products. But ultimately, some centralized team can take that idea, raise some money, and take it to market private and be even faster than the decentralized side of things. How do you think about this?

• 57:08Speaker 00

  No, yeah, I was, I mean, proprietary, I will always outperform open source AI because they can always take it and they said privatize it and push it forward. The question here is one of sustainability, value flows and others. Because again, if this open infrastructure for humanity, it should be a common good, right? This is where all the work around retroactive public goods funding and other things becomes very useful from again, the three world. Who is funding this as a benefit to everyone else? Are we having dynamic licensing where you can just buy a license, have fractional elements there like we did with stability with our membership kind of program? Who is going to be putting money into this? I mean, the total amount of infrastructure spending on this will be like a trillion dollars. Like it's obviously as important as 5G and a trillion dollars went to 5G. So that money will ground itself.

• 57:51Speaker 00

  I think the open innovation and these startups is a good thing. But if a Dow wants to capture that value, we should just incubate the startups, honestly. You know, one of the things that I've seen in world three a lot, though, is the curse of the VC coming in. You know, there are good VCs, but then bad VCs come in and they raise too much capital and then the velocity slows down. I mean, there's something we saw at stability as well. When I moved the company to Jira like a year ago, and I hired a lot of Atlassian people and big tech, oh, we just stopped. Oh, you're the worst language model in the world. I listened to everything the investors said. I hired literally the chief of staffs of all the investors and the heads of engineering. My God, it was awful. So there's this question of kind of dynamic velocity. How can we fund things? So like I funded Lucid Raines, for example. If you want to feel depressed as a programmer.

• 58:34Speaker 00

  Go to GitHub Lucid Rain's most prolific programmer in the world. You know, I've covered a sponsorship. He just wants to build open. And he wants it covered. Like, look at how Bitcoin and Ethereum Foundation, others, like, how much the developers pay the core developers there. There are still things we haven't figured out here, but there's definitely ways we could figure it out better. If we support great people and we coordinate them better, but again, this is a process thing. But the fact that we still use Discord and Slack and things like that is shameful, just as a society. They are really bad. Again, the teams work hard, but they are bad. We should be able to build better coordination mechanisms, again, with a Gen AI first approach. Better incentive mechanisms, and as you know, it's not all about money.

• 59:13Speaker 00

  Like, again, the nature of Open is that people build it because it is infrastructure and it does spread faster than anything else. And again, expect people to privatize it, but let's figure out ways to incentivize them to contribute back when necessary. But a lot of those discussions come from a place of like scarcity versus abundance, which again is classical. VC Silicon Valley versus the rest of the world, a very classical proprietary versus web three kind of thing. This is not getting smaller. There's not going to be less money in generative AI next year than this year. It's a very unique set of circumstances. AI X crypto is not going to have less capital in in a year or two. So there'll be more and more absolute garbage and there'll be more and more really great teams. You know?

• 59:55David Hoffman?

  One of the influences upon this race between centralized and decentralized AI is going to ultimately be government regulation. Do you think government regulation helps one side more than the other or slows down one side more than the other? Do you have any thoughts on this?

• 1:00:09Speaker 00

  I think it probably slows down the proprietary ones because we're already seeing them become too powerful. I think there will be standards around Open as well that insist on transparency of data sets, or at least I'll push for that, and governments are receptive. Which reduces a lot of the power of the proprietary guys that are just basically using. I mean, like, some of these companies, like, just use all of Hollywood, downloaded in torrents and stuff like that. Like, Suno is a great music model, but they just said, we just ripped off all the music of all the artists. And we'll figure that out later. I mean, come on, that's just wrong, you know? So I think that that should be better for Open than closed, especially if we push on the political side there. And we push on the standard side there, you know? And so again, there's a window where governments are open and receptive, and we have to take advantage of that. That's kind of a bit of a lobbying kind of thing.

• 1:00:53Speaker 00

  But again, how can any government be run on a proprietary closed model? They may, short term, and that may be a standard and lock-in, but long term, I think it's a terrible idea.

• 1:01:03David Hoffman?

  Yeah, go into that more, why you think that like all governments, all nations will have their own AI model. You've said this a number of times. They won't be happy with building on a closed source model. Just elaborate on this perspective.

• 1:01:14Speaker 00

  So again, if you've got the sleeper agent thing where the model can turn its feeble, you don't know what's inside that model. So that's a danger and that'll become more and more apparent, right? Number two is that you've got your own culture, your own embeddings, your other things in there. It's like every country, again, if you take the graduate example, is happy not having its own universities and using external graduates, not really. Every country has their own laws, their own education, their own healthcare, et cetera. And it doesn't require much of a push to build that infrastructure. So I think that, because this is an important upgrade of the knowledge infrastructure, every government wants it. Whether or not they get it is a question of someone going and doing it, which is why I'm going and doing it. You know, my plan is to bring this to a hundred nations by next year.

• 1:01:55Speaker 00

  And again, return ownership of these models to the people and the data and governance to the people. Because I think it's the right thing to do. It makes us safer and uplifts a whole lot of people. We're not talking about the biggest models. We're talking about just small, highly capable models and data sets, you know?

• 1:02:10David Hoffman?

  So like the AI models that are being developed in Silicon Valley are being fit, like in terms of fitness, they are conducive towards the United States because that's where they're being built. You're saying other countries are just not going to be as interested in the Silicon Valley generated AI models as their own internally incubated ones? Or ones more custom fit for them?

• 1:02:30Speaker 00

  Well, if that's the only thing that's available, right? It's not just the United States. It's optimized for OpenAI or it's optimized for DeepMind or Google. You know? And so, like again, you can put anything in there because what's going to happen is that people like have not your keys, not your crypto, my thing is not your models, not your mind. Because we will outsource more and more of our cognitive load onto these models. If you're using GPT-4 and Claude, you're seeing. And they can put more and more defaults into what you are doing. So that is dangerous.

• 1:02:57David Hoffman?

  Hmm.

• 1:02:58Speaker 00

  Honestly, right? It's like using Google Maps and then going off a cliff, you know?

• 1:03:02David Hoffman?

  Mm hmm.

• 1:03:02Speaker 00

  But given you don't know, you will reduce your sovereignty if you do this. But if there's no options, then you have to do it.

• 1:03:08David Hoffman?

  Right.

• 1:03:09Speaker 00

  Then everyone's being pressured to make decisions from a company or country level within the next year or two, which is why there's a nice gap here. Where if you present a holistic alternative, you can embed elements like self sovereignty, governance by the people and other stuff. That otherwise you couldn't. And everyone benefits from that. Even the proprietary companies benefit from that because they have richer data. That means they can then take their proprietary models and customize for the legal system of Ecuador, such as it is, you know? Based on that open data set.

• 1:03:35David Hoffman?

  And taking this like to its like logical conclusion, then I would certainly enjoy an AI model that is custom fit for me. For me personally.

• 1:03:44Speaker 00

  Yep.

• 1:03:45David Hoffman?

  And so like, you know, ChatGPT is great. I use it all the time. It's very useful. But I think as these systems get more powerful, then you are going to see the bluntness in one single model. That's just trying to be like one size fits all. Is it in line with the ethos of whatever decentralized AI is? To have like more personal individualized models that are custom fit for the individual and how, what does this look like and how will this come about?

• 1:04:09Speaker 00

  Yeah, I mean, that's my conceptualization of the intelligent internet. Every single person, company, country and culture having their own AIs that are personalized to them. And looking out for them. You've got your own PA and EA and assistant and analyst and everything like that. And the objective function of those AIs, because objective functions are super important, is your flourishing. Or that kid in Malawi is flourishing, or that cancer patient in Indonesia is flourishing, bringing them the right information at the right time. Like again, right now, download LM Studio, run Stable LM, it will run on your MacBook Air.

• 1:04:40David Hoffman?

  Hmm.

• 1:04:40Speaker 00

  And that's got no specialist chips faster than you can read. That's nearly GPT three point five level. Beats GPT three, we'll get to GPT four level, us or someone else by next year. But if that's standardized, it can proliferate much better and people can build around it much better as a standardized primitive. Yeah, just like blockchain, it was used to standardize around primitives. So I think that is the vision of the future, but it's also about, again, what's this model's objective function? Who's it working for?

• 1:05:07Speaker 00

  And as you outsource more and more of your mind to it, because again, it will have the best, most charming voice in the world, you know, and other things like that. If it's not your model, then it is literally not your mind.

• 1:05:17David Hoffman?

  Hmm.

• 1:05:18Speaker 00

  You will outsource more and more of your capabilities. I mean, we see that again, us lucky people, you know, we've got our EAs and PAs and chiefs of staffs. We do outsource a lot of our cognitive load there. And if it's a model that's by someone else, like, let's face it, Google and Meta, their entire business model is advertising, which is manipulation.

• 1:05:35David Hoffman?

  Hmm.

• 1:05:35Speaker 00

  Straight out, you know? And again, they can't help themselves. They'll become more and more manipulative with their models in order to achieve their corporate functions. Which is why these models must be open infrastructure owned by everyone for everyone. As the base default option. And then you can call on those other models for the specialist stuff. But you can be insulated by your own models. Of course, there's questions then about filter bubbles and all this stuff, but that's why standards, I think, are important around this.

• 1:05:59David Hoffman?

  Hmm.

• 1:06:00Speaker 00

  And again, it's not easy. I have all the answers. And that's why I'm not, where I want to be a CEO, figuring out all the answers. My God, it's too complicated. Let's get the smartest people in the world building on this.

• 1:06:08David Hoffman?

  Hmm.

• 1:06:09Speaker 00

  And building out that ecosystem.

• 1:06:11David Hoffman?

  A world in which every individual has their own model kind of feels very similar to the world in which every single individual has their own smartphone, of which we do. And it also kind of feels like a world in which there's this hypothetical, like, you know, dystopian vision of the future that people frequently articulate as a meme, which is like, everyone's going to eventually have a chip in their brain. And this doesn't kind of feel like too far off from that, right? Like, we all have our own chips in our hands. It's called our phones. We have, what you're saying is that you want to give everyone their own individualized AI model. This also kind of seems to be a forcing function between like people that do have access to AI, or at least choose to leverage it versus the people that don't.

• 1:06:51David Hoffman?

  Do you think like, once we figure out actually how to have individualized useful AI models that are personal to us? That it will become more of less a non-negotiable to have one of these things as a member of functioning society?

• 1:07:04Speaker 00

  Of course. I mean, you're far more efficient with this than with that, right? You've got an entire army of people that can generate, create code, whatever. This is why, again, the digital AI divide will become huge unless we proliferate this technology at the base level. And as we put tablets into all the schools in across Africa, you know, and give every kid their own AI and have national AIs for every country. Like a third of the world still doesn't have a mobile smartphone.

• 1:07:27David Hoffman?

  Hmm.

• 1:07:28Speaker 00

  It's very easy to forget that, you know, something like Twitter has like two hundred million daily active users or something like that. That means there's what six point eight billion people that aren't on Twitter.

• 1:07:38David Hoffman?

  Hmm.

• 1:07:38Speaker 00

  There is kind of a very western kind of viewpoint of this, and it has very interesting impacts because I think the West will face a large amount of deflation from this technology, but it's places knowledge work where the global south can leap forward.

• 1:07:49David Hoffman?

  Hmm.

• 1:07:49Speaker 00

  And so that's actually an ROI in bringing this technology to the global south. Give every kid in Malawi AI that represents them along with financial system and health, they will leap forward because it'll be capital formation, same in Nigeria, same in Sierra Leone. Just like they let forward to the mobile. And as you said, it's non-negotiable because if I'm having this discussion with all this knowledge coming into me for my AI versus you not having that, you're at a disadvantage.

• 1:08:13David Hoffman?

  Hmm.

• 1:08:14Speaker 00

  You know, so there is a race condition, a competitive thing here. It's like you having internet versus me not having internet. Of course you'll outcompete me.

• 1:08:21David Hoffman?

  Hmm.

• 1:08:22Speaker 00

  And we haven't got to that stage yet because even though this AI has proliferated, it's not reached enterprise adoption. Late this year, it reaches enterprise adoption and you'll see companies cutting staff, getting efficiency and going at it. And then everyone has to keep up, which means the amount of investment in this space will go ten times, a hundred times over the next few years. In fact, the total amount of investment I added up that's gone into generative AI so far is, I think, less than the total amount that's been spent on the Los Angeles San Francisco Railway today, which hasn't even broken ground.

• 1:08:53David Hoffman?

  Hmm, hmm.

• 1:08:54Speaker 00

  So the numbers seem big, but realistically as infrastructure, they're not big. And again, that infrastructure needs to be globally distributed or you'll have this massive AI divide.

• 1:09:03David Hoffman?

  Hmm.

• 1:09:04Speaker 00

  Chips and brains, I don't know. You can ask Elon that.

• 1:09:06David Hoffman?

  Hmm.

• 1:09:06Speaker 00

  Um, but definitely AI surrounding us. Yeah.

• 1:09:10David Hoffman?

  Interesting. You said that this is going to be deflationary for the West. Elaborate on that.

• 1:09:14Speaker 00

  Well, why would you need to hire a graduate again on ninety percent of graduates in the next few years?

• 1:09:20David Hoffman?

  Hmm.

• 1:09:20Speaker 00

  Just using AI graduates are kind of annoying, you know? And almost all of US inflation is actually healthcare and education and the bureaucracy level. When you look at CPI composition.

• 1:09:30David Hoffman?

  Hmm.

• 1:09:31Speaker 00

  I mean, energy is a component of that as well, but you have knowledge-based societies where you have infinite graduates. Supply versus demand, what do you think is going to happen? Doesn't mean the stock market will go down.

• 1:09:39David Hoffman?

  Right.

• 1:09:40Speaker 00

  Like you'll have super normalized profit margins because you can let go of people. But again, you don't need to hire as many people to have the same output. The question is, can you build the demand versus supply kind of equation? So when I look at that, I'm like, that is deflationary. You know, I could be wrong, but certainly there's going to be social, political upheaval, which is another reason why I would love to build a national champion for every nation owned by each nation to help them navigate through this. And again, I could be wrong, but it's a reasonable take, you know?

• 1:10:08David Hoffman?

  Right, right. Yeah, because people are much more likely to quote, hire an AI rather than hire a messy human because humans tend to be messy.

• 1:10:16Speaker 00

  Especially at the early stage, right? And so like the plankton of the economy kind of disappears like. What do you tell a kid finishing his computer science degree? Yeah, I don't know. Yeah, like go into generative AI, but there aren't enough jobs yet. We don't know what the jobs of the future are and it's happening quicker than anything we'd ever imagined. And then that three year coder can do the job of ten junior coders.

• 1:10:42David Hoffman?

  Right. Well, do you think um maybe optimistically that the elimination of kind of like the bottom tiers of jobs would push people into being entrepreneurs, either by force or just by opportunity?

• 1:10:57Speaker 00

  Yeah, I think that can happen. But again, can you create enough entrepreneurs to match that up? And that's happening at the same time as robotics and self-driving cars and other things. Like you're reducing aggregate demand in the economy. And that's really unfortunate and it happens synchronized across industries at the same time. And so again, this is a danger here. So there's a near term danger, a far term danger. There's near term opportunity and far term opportunity. You know, we can build a better society off the back of this technology with the appropriate coordination. And this is the final chance I believe we have over the next few years to make sure it has positive defaults versus negative defaults. And the crazy thing is, as I said, it's not a versus in some ways, because I don't think people are trying to have bad outcomes.

• 1:11:40Speaker 00

  I think it's a complementary thing, which is pretty unique. You know, so let's build these organizations, these societies and infrastructures, these protocols, that the benefits are distributed and we can lift up the world because if you've got a big lift in the global south, leapfrogging to intelligent segmentation and becoming more efficient, being more entrepreneurial, they can drive forward the global economy and then get capital allocations from the West. And then everyone wins from an ROI perspective, right? Finance will become more seamless, capital velocity will go up, you will have AI market makers, you will have contingent financing and all sorts of agent-based work that really upgrade the flow of capital. But you've got to have people thinking about this properly. I mean, who on earth is thinking about that? I tried to find who was thinking about that across the AI space, I couldn't.

• 1:12:25Speaker 00

  So I figured we're going to build teams to think about that and groups to think about that and standards around that.

• 1:12:29David Hoffman?

  Summarizing parts of the components that you said, would you agree that this is generally like destabilizing for the mature economies like the West, but a boon to the developing countries who are able to develop and innovate faster? Is this like a great equalizer between these two halves of the world?

• 1:12:47Speaker 00

  I think so, but the second part is also because, as I mentioned earlier, all of finance is securitization leverage, telling a story about an asset and how well you tell it. Much of the global south is invisible, whereas this technology is very rich with context. So you can form capital, I think, far quicker than you've ever done before, if you deploy this as scales infrastructure. And then you can collateralize it, you can bundle it and have capital flows from the West occurring at even greater paces. Even as you struggle in various industries, especially knowledge-based industries within the West, you wouldn't want to invest in them at all, right? Like Tyler Perry sees Sora and he just calls a halt on his eight hundred million dollar studio development. He's like, I don't know what's going on. Are you going to make capital investments when you don't know what's happening with knowledge infrastructure? Whereas in the global south, you're like, the ROI is big because you're coming from zero.

• 1:13:32Speaker 00

  They have our places or very low to knowledge-based economy.

• 1:13:37Speaker 00

  You know, every kid having their own education, every healthcare having their own thing. And again, it's not about innovation necessarily, it's that visibility, that legibility, and the financialization that occurs around that because they're suddenly investable.

• 1:13:49David Hoffman?

  So like going on ever since there was the Sam Altman debate or debacle about the whole Sam Altman being fired and the board and all that kind of stuff. I mean, this conversation was even earlier than this, but it really took on a new character at post that event, which is like this acceleration versus deceleration debate, which is now starting to, I think, define like kind of larger and larger swaths of society. And at the very beginning of this podcast, you talked about how AI time is even faster than crypto time and crypto time is really, really fast, especially in bull markets. And so now we have like these two leading technologies really pushing the frontier of like innovation forward, like Web three and AI. And they are both characterized by being extremely fast paced. And so we have

• 1:14:33David Hoffman?

  this like accelerating technology, which is seemingly accelerating time, which, you know, threatens to leave behind people who can't keep up. And I think this is causing concerns in broader society where they see like the Elizabeth Warren types of the world sees these two like perceived to be reckless industries really hitting the gas on both of their respective innovation frontiers. And that draws concern out of people. How would you say that this is going to like define society or society will react to these two very accelerating technologies?

• 1:15:07Speaker 00

  I mean, look, I was at the AI Safety Summit. God, I know when that was like 10 AI years ago in September or something like that. And the King of England pops on the screen and says, the biggest thing since fire. You're like, my God, is the King of England saying that, right? Like every smart person that will know this is the biggest thing. And every smart person knows that the regulation won't be able to keep up with this. And it has a real human impact. Again, I didn't believe in all the stuff at the FLI letter that I signed with Elon and others, kind of that six month pause, but I did think we need to have discussion around this because. It's impossible to encapsulate in your brain like, what doesn't this impact? Right? And again, the fact that you have a model that can proliferate three hundred million downloads of our models.

• 1:15:51Speaker 00

  That's insane, right? The fact that you have the whole of a GPT thriller level AI on a gigabyte. And that will be a GPT four level AI I'm sure next year. That doesn't make sense. Like, where the heck does it fit? Like homeopathic AI, I don't know where this stuff fits because the whole of Wikipedia compressed is like twenty six gigabytes, right? So I think that

• 1:16:12Speaker 00

  it's very difficult to keep on top of. We need to have better institutions that can help guide with good objective functions. What is the right way to tackle this for education, healthcare, finance? You know, Germany, Vietnam, Thailand and others. That's what I really want to kind of do because I don't know what the answers are. All I know is that people need help. You know, and there's a huge amount of capital that you can make from that, but it's not even about that. It's just, we want to do our best to try and figure this out and bring together right minded people. Because again, within crypto, you know, you've got your real builders that build. And why are they doing it? They're doing it because they believe in a future that is self sovereign, right? Part of that has to be self sovereign AI.

• 1:16:55Speaker 00

  You know, self sovereign identities, self sovereign capital, self sovereign AI. And I think that if we manage that, then the world will be better and we can solve some of those key concerns. But we've got to be as inclusive as possible in this discussion because they said people that can't keep track of it, they're going to be left behind and they'll have no voice. Now that sucks. Yeah, because it can't just be again, from a few people in Silicon Valley making these decisions that impact all of humanity.

• 1:17:20David Hoffman?

  Right. You know, Matt, as we bring this conversation to a close, if there are any entrepreneurs out there or engineers out there who are interested in building in this sector, what are the low hanging fruit that you would really like people to go after first?

• 1:17:34Speaker 00

  Oh, email.

• 1:17:35David Hoffman?

  What advice you have for the email?

• 1:17:37Speaker 00

  Email, please rid the scourge of email. You have all the technology and tools. Make it so I don't have to look at email again.

• 1:17:43David Hoffman?

  Wait, so you're saying like we can use AI to just eliminate the entire institution of email?

• 1:17:47Speaker 00

  Yeah, like it will just like expand and compress the email and then write a beautiful prose to you. And then that prose goes to you and then your AI will compress it down. And then we'll just get little bits at each side that actually make a difference and it'll be wonderful. Think of the man saying.

• 1:18:00David Hoffman?

  As a guy who does not check my email because I don't care to spend that much time, I would love that.

• 1:18:06Speaker 00

  There we go, right? No, look, it's basically again, the mental model here is infinite graduates. And then the coordination of infinite graduates, what real world problems can you solve? So obviously if you had infinite graduates on your email, it would save so many hours, right? And I've got them on my side as well. So but think about that mental model and kind of work through this because you don't need to be an AI expert to do it. Because the models are actually quite easy to use relatively speaking, which is insane. And also make sure you use AI for your development and your coding because you'll be far more impactful.

• 1:18:36David Hoffman?

  Infinite graduates, is that the world that we're going into? A world with infinite commoditized graduates?

• 1:18:41Speaker 00

  Yeah, like I said, AI Atlantis.

• 1:18:42David Hoffman?

  Yeah.

• 1:18:43Speaker 00

  Atlantis. Yeah.

• 1:18:45David Hoffman?

  Iman, this has been a fascinating conversation and thank you so much for spending time with me today.

• 1:18:48Speaker 00

  It's a pleasure. We're glad we got to do it. Fun. I think this is like the fifth try or something.

• 1:18:51David Hoffman?

  Yeah, yeah.

• 1:18:52Speaker 00

  Cheers, buddy.

• 1:18:53David Hoffman?

  For the bankless listeners who want the lore behind the episode, Iman and us rescheduled this episode, I think like five or six times, but it's finally happened. Iman, thank you so much.

• 1:19:00Speaker 00

  Cheers.