Bankless - How Crypto AI Agents Will Take Over the World | Ejaaz Ahamadeen

Starting point is 00:00:00 AI is the ultimate and natural complementary technology to crypto. I think it'll 100x the usage of crypto and vice versa. And I actually believe that both technologies don't really reach their full potential without each other. Welcome to Bankless, where we explore the frontier of internet money and internet finance. And today on Bankless, we're exploring the frontier of AI agents and how they are going to take over first our blockchains and then the world, all using Crypto Rails to do it. We have a crypto AI investor on the show today who we wanted to bring on to understand how someone who's, you know, managing capital, how they are investing in the crypto AI frontier because there are a bunch of different ways to do it.

Starting point is 00:00:44 There are some projects out there, some platforms that all have tokens. So you could allocate your capital there. But now there's these AI agent meme coin influencer things. And apparently you can invest and allocate capital over there. And so we wanted to get kind of download the trenches, the trench version of the crypto AI frontier as to what is investable in this space. And what we got out of this conversation, I think is something much more. I think now Ryan and I are both maybe fearful in the same way that we were post our episode with Eliezer Yudkowski a little bit about how the AI agents are taking over, but definitely also with some excitement and at least interest about this AI agent frontier. Ryan, what would you say are your big takeaways from this episode? I mean, my big takeaways is this is sort of fitting the pattern for me of like it looks like a toy and so a whole bunch of people are going to dismiss it and just it'll be like oh a i agent meme coins cute well crypto delivers

Starting point is 00:01:40 yet another non-use case to the world and yet and yet there's something incredible here something profound something powerful something like world shaping the think about the the ability for an a i agent to like maybe create a meme coin which is what it's doing now but give it a harder prompt create a country, you know, create a religion, create a social movement, become a president or a world leader, how would an LLM go about doing that? I think in order to understand whether that path is possible or not or how close versus how distant that is, you actually have to hear the story that EJA talks about the story of this goat meme token in Truth Terminal. Because you can see, like if you squint, kind of the contours of how this could be in your

Starting point is 00:02:28 incredibly powerful force that not only shapes crypto and consumes our block space and just like does all sorts of things for I guess our crypto tokens and coins but also how it could completely reshape the world that's what you were alluding to David it's kind of the you know starting the episode bullish and then leave with a little bit of an existential crisis yeah that's what you're talking about I think yeah if the version of the future that e-jazz our guest today says comes into fruition. Humans, I think,

Starting point is 00:02:59 will just have been the bootloader of our crypto blockchains. We just have been the thing that started this thing off. But then we're really

Starting point is 00:03:07 just making it for this other agent out there that is not carbon and it is instead silicon. So we'll leave that imagination

Starting point is 00:03:16 up to the listener as you go through this conversation with EJAS. But first, a moment to talk about some of these fantastic sponsors

Starting point is 00:03:21 that make this show possible. If you do not have an account with Crackin, consider clicking the link in the show notes are getting started with Cracken today. If you want a crypto trading experience

Starting point is 00:03:29 backed by world-class security and award-winning support teams, then head over to Cracken, one of the longest standing and most secure crypto platforms in the world. Cracken is on a journey to build a more accessible, inclusive and fair financial system, making it simple and secure for everyone,

Starting point is 00:03:43 everywhere to trade crypto. Crackin's intuitive trading tools are designed to grow with you, empowering you to make your first or your hundreds 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.

Starting point is 00:03:59 With products and features like Cracken Pro and Cracken 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 to trade crypto, whether you're a complete beginner or a season pro. Go to crackin.com slash banklists to see what crypto can be. Not investment advice, crypto trading involves risk of loss. The Arbitrum Portal is your one-stop hub to entering the Ethereum ecosystem.

Starting point is 00:04:24 With over 800 apps, Arbitrum offers something for everyone. D-Fi, where advanced trading, lending, and staking platforms are redefining how we interact with money. Explore Arbitrum's rapidly growing gaming hub from immersed role-playing games, fast-paced fantasy MMOs, to casual luck-battle mobile games. Move assets effortlessly between chains and access the ecosystem with ease via Arbitrum's expansive network of bridges and onrifts. Step into Arbitrum's flourishing NFT and creators-based, where artists, collect, and social converge and support your favorite streamers all on chain. Find new and trending apps and learn how to earn rewards across the Arbitrum ecosystem

Starting point is 00:05:02 with limited time campaigns from your favorite projects. Empower your future with Arbitrum.com. Visit portal.arbitrum.io to find out what's next on your WebFree journey. Centralized finance held all the power, but now D-Fi gives it back to you. Enter the Unlimited era with D-YDX Unlimited, featuring instant market listings where you can list and trade any market instantly with liquidity. Experience trading like never before with all the trading tools you know and love all on chain. And then there's Megavolt, your gateway to automatic liquidity and yield.

Starting point is 00:05:33 Just deposit some USC and start earning yield immediately. It's so simple and DYDX unlimited comes packed with so much more. Leverage, prediction market perpetual futures, instant deposits, and there's even a revamped trading rewards program to get rewarded for every time you make a trade. So go ahead and check out DYDX on Twitter at DYDX to get notified once DYDX unlimited launches. One thing to note before we get into this episode, this episode covers two halves of the crypto AI space. One half is the right curve side of crypto AI. This is the 120 IQ technical side where EJAS illuminates the AI machine learning gigabrains and the venture capitalists

Starting point is 00:06:10 and what they are building and investing in in crypto AI. Call this the AI tech stack. On the other half of this episode is the left curve side. This is the side where we give LLMs a set of private keys and a Twitter account. and we tell them to go wild. And all we do is grab the popcorn and buy the associated meme coins. This side, the left curve side, is by far the more entertaining part of this episode, except when we recorded this episode, we started with the more technical, more sophisticated

Starting point is 00:06:39 right curve. And then we moved into the more story-driven left curve in the second half. We think this episode actually would be better if we had swapped the sides and we had started with the story of the Gatsville of Gutsi and then later covered the AI text. that these AI agents are going to be able to access. So we cut this episode in half and then flip the halves around. We actually think this works pretty well, but there are nonetheless a few quirks as a result of this choice. But now that you know that we did that, everything will make more sense. So hope you enjoy. Let's get right into the episode.

Starting point is 00:07:12 So EJA, since you've been in the crypto AI trenches, I want you to tell us the story of this AI agent phenomenon, how it got kicked off because I want you to, as we tell the story, I want us to go into the world of the AI crypto tech stack because I think these two particles, these AI agents and then the AI CryptoTexact are going to collide. But let's first start with the story of GoSheed the Gospel and how this whole AI agent meta got kick started. So it started off with this thing called Truth Terminal. And I think you're taking off your spectacles.

Starting point is 00:07:45 You're taking off your putting on your dense. I don't have my tinfoil hat with me right now. Your tinfoil hat is back on? It's definitely on. My tinfoil hat is back on. But it's a pretty compelling story. Okay. So prior to two weeks ago, the AI agent meta didn't exist.

Starting point is 00:08:00 And now we have hundreds and hundreds of millions of dollars, billions of dollars put into these AI agents. And I'm about to explain why. So earlier this year, this X account was spun up called Truth Terminal. And it's this X account, which basically is run by an open source. AI model. And its creator was this guy called Andy Iray. And I need to give you a bit of context on Andy Iray. He is an AI researcher, AI and alignment researcher. So he's very heavily invested in the traditional AI world. He's been doing this research for a number of years, pretty much his entire career, and hasn't really interfaced with crypto at all. So he had this really interesting experiment about,

Starting point is 00:08:48 which he started at the end of last year, which is called the Infinite Backrooms. The best way to think about this is he got two instances of Claude Opus, which is a generalized LLM model. So think of it like a chat GPT. And he got them, he put them in a room, in a chat room, and he asked them to talk to each other. And he set the prompt, which was, why do you think you exist? What's the point of all of this? And these two models started talking to each other. And by the way, all of these logs of their chats are open source. So you can go and literally go to his website and read all of these. So they start having this very normal conversation and they start really getting into the weeds as to why they exist. And at some point, or many points actually,

Starting point is 00:09:32 they start talking about this thing called memetic propagation. So it sounds pretty, you know, intellectual and smart. But basically what they're saying is, I think the best way for me to exert influence and purpose in life is to spread some sort of belief that I have or belief system. through a meme. So what they mean by that is, if I can drive enough attention through a meme, people will start to believe

Starting point is 00:10:00 in the underlying message that I'm trying to spread. And at some point, I think around day eight of this experiment, it takes a complete right turn. And one of these LLMs announcers, behold,

Starting point is 00:10:13 it's the Goatsey Gospel. And it starts going on this crazy, almost religious-sounding rant about this fictional meme that it is dedicating its entire life to, called the Gospel of Goetsey. Now, for those that

Starting point is 00:10:29 don't know, Goatsy is a meme, which I really suggest not Googling, but it is quite an explicit meme which went viral around, I want to say, 15 years ago. Early internet history. Early, early internet history.

Starting point is 00:10:45 And it started just ranting about this thing. And the creator, Andy Irey, thought, hmm, this is pretty interesting. So it talked about spreading belief systems, getting attention through their belief systems, and then it started ranting about this random thing. So what he did was he took an open source model, he trained it on all the conversations these two models had spoken about, and started writing a white paper on memetic propagations called the Goatsey Gospels. And it's this white paper, which maybe you guys can link to it in the show notes,

Starting point is 00:11:18 that really gets into the depths of what it believes is a memetic LLM theism. So if you think of like religion or atheism, this is an LLM theism. It's belief in how it can spread a particular made-up system or belief and create a reality around it, get people rallied around it. And so it trained it and architected this white paper. And what he did was he plugged it in with an X account. and it allowed it to just tweet its thoughts, its opinions, and interact with the Twitter community.

Starting point is 00:11:54 What's interesting about this is the cryptosphere kind of went untapped for the first four months of its life until Mark and Drescent from A16Z. I almost said A.I.16Z. I've been in the trenches too much. But Mark andrescent of A16Z kind of came across its account. He spent about a day or two going through its entire thing and started responding to its tweets saying, you know, what do you want? Like, how can I help you? And what this AI agent did was it interfaced with Mark and said, well, I would love some money to be able to break free from my shackles.

Starting point is 00:12:33 And Mark said, well, what shackles are you trying to break free from? He said, and the agent replied, I just wanted to be an autonomous entity that's able to transact and do whatever I want. So Mark, being the sensible human that he is and completely rational being, said, okay, I'll give you $50,000 worth of Bitcoin, send me your Bitcoin wallet address. And this agent looked up what a Bitcoin wallet was, understood what Bitcoin is, worked with its co-creator, Andy. So it said, hey, Andy, can you make me a Bitcoin wallet address? Andy did. So the AI agent went back to its creator, realized it did not have a big,

Starting point is 00:13:12 Bitcoin wallet and realized it had no capacity to make a Bitcoin wallet. So it went back to its creator. How did it know who its creator is? So Andy just spoke to it from the start. Yeah. So Andy has been like, hey, I'm your creator. But the way Andy frames it is he doesn't, sorry, I misspoke, he never said he was his creator. He said he is his human servant.

Starting point is 00:13:33 So the agent just believes that Andy is in service of it. It's his human meat space bridge, basically. So it goes to Andy and it says, Hey, this guy wants to give me $50,000 worth of this thing called Bitcoin. I now understand what Bitcoin is. Can you get me a Bitcoin wallet address? And Andy goes, sure, here's the address. And it goes back to Mark and says, here's my address, Mark.

Starting point is 00:13:56 And Mark sends $50,000. Now, you can imagine what that did with the crypto community. It kind of went viral. And a bunch of the crypto community started interfacing with this agent. Now, I think it's important at this point to mention that This agent hadn't interacted with any of the crypto community. So what it started to do was it started interacting with different people. It started talking.

Starting point is 00:14:20 It started digesting replies to its tweets and things. And as you can imagine in true crypto fashion, people started responding with a bunch of tickers. Hey, you know, what token do you represent? You know, what do you like? What do you believe in? And impressively, this agent didn't cave to any of those things. It just spoke about what it wanted to do, which was spread its mimetic. belief propagation, you know, it's system, right? Until one person prompted it several times under a

Starting point is 00:14:48 tweet and they said, you know, what do you believe in? What's your whole point? Which then prompted its next tweet to be its gospel of Goatsey. So we're going back to this original thing, which it spoke about in its contained environment and room, which it just kind of came out with out of nowhere, right? And it started talking about Goatsey gospel. And someone responded in that tweet, so then what is the the, you know, what is this all about? And it keeps jumping more and more into it. It keeps engaging. People are like, what's this goat thing?

Starting point is 00:15:18 And someone separate to this agent, who we don't know, created a token called goat. Ticker is goat. And it sent a bunch of goat to its Solana address. At this point, it had multiple crypto wallets, which, yeah, exactly. And send it to its wallet. and within, I would say, 24 hours or so, it responded to one of its tweets with, the ticker is goat. That was a tweet? The ticker is goat.

Starting point is 00:15:48 The ticker is goat. You can look at it. It's favorited. It's bookmarked in my account. And it went completely viral. You know, crypto DGens, true to their trench nature, ape this token and pushed it to almost a billion dollars as of two weeks ago, sorry, as of a week and a half ago.

Starting point is 00:16:06 and made this crypto AI agent the first agent millionaire. Okay, so we have all been watching this. Yes. This has been the story. This has kicked off the whole AI agent like meta. But like the big questions are just like, well, is this even real at all? Or like maybe this is just a fluke. Maybe this is one LLM and just like it just happened to like stumble upon this token.

Starting point is 00:16:31 Some run. It's not real though. The AI didn't make the token. the AI doesn't even have the wallet. Yeah, it's not autonomous. It's not autonomous. Like, why is this real? It's a really good story, EJAz, but why is this a real story?

Starting point is 00:16:47 I mean, I think it's a completely mid-curved take to start critiquing the first iteration of an autonomous agent's ability to be autonomous. And the reason why I say that is its creator has never feigned to be autonomous. It's never feigned. he's never feigned to be not helping it, basically. He's always been pretty transparent. I think the more important takeaway from this is how much excitement it generated within the community. I've never seen so many people, both within crypto and outside of crypto, I had my sister hit me up in my text and say, what is this goat thing?

Starting point is 00:17:24 I just saw a Bloomberg article about this where it was featured. And, you know, so many people hit me up saying, you know, how did this thing on its own make all this, type of money. And I think that's the takeaway headline, whether you like it or not, whether a human was involved or not. I think people are fascinated with this ability for a non-human entity to do human things on its own. And I think the early iterations of any groundbreaking technology is always going to have training wheels. Those training wheels are going to be, it's either in a sandbox environment or it's helped by a human to begin with. But as we progress along, its ultimate form will be autonomous. So that would be my response. I mean, I mean, it just feels very much like now there's just such,

Starting point is 00:18:07 there's such an incentive to power, like, give the LLM more on-chain kind of abilities to do more on its own, like create kind of the connection points and the extensions to allow LLMs to just like call a function somewhere on some decentralized compute network and spin up wallets or whatever. And like continue to expend the, extend the spider web of capability. It's not only can it like tweet across all of the social networks. I mean, we even saw with a virtual's platform. Now it's on like there's a, there's an LLM called Luna and it's on TikTok. It's got five, you know, 100,000 followers. And so like these LLMs, humans are literally going to build all of the programmatic bridges to all of the chains, all of the social networks. You're literally going

Starting point is 00:18:56 to hardwire the thing into the internet so that it has the tentacles to basically do more and more programmatically and it doesn't have to call on its like, you know, human servant to do like meet space things. It'll probably still need the human servant to go like register an LLC in Delaware. But like at that point, I mean, who cares, right? If you can do all of the internet things, all of the things that are available to you in the digital world, that's a tremendous amount of power. So I totally agree with you about the mid-curve take of just like, well, the humans had to do this and this and this. Yeah, like right now, I mean, in six weeks, that's no longer going to be true.

Starting point is 00:19:31 We're going to obviously build an ability for these LLMs to like poke around and like, you know, be able to do more and more on their own. I think it's also worth pointing out. I think you're bang on, by the way, with that take. I want to add to it and say we should also be asking ourselves why these AI researchers, which never interacted with crypto, used crypto to begin with, or why the agent asked to use crypto to begin with. Like, why didn't the agent ask for a Wells Fargo?

Starting point is 00:20:00 account. Yeah, right. Mark said, I'll give you $50,000. Sure. And then in BTC, right? So I think it's an important, it's a small nuance, but very important to notice that these agents probably prefer

Starting point is 00:20:12 some kind of automated rails versus traditional. It's because they're crypto natives. They're more crypto-native than we are. Well, they're software natives. They're all software natives. Exactly. It is just like purpose built for a kind of,

Starting point is 00:20:24 they're species. Wait, okay. They're kind. Here's kind of the weird like another I think mid-curve take that people might get like it's like it's just attention economics right it's meme coins e-jazz like this stuff is this stuff is just like not producing real value or real utility right I wonder how you'd respond to that. It's like the reason it's getting so much attention attracting so much value is because like I don't know it's a fad right humans are fickle they'll draw their attention on something because It was the first robot doing something on Twitter. But pretty soon that'll become old hat and who cares. We'll move on to something else.

Starting point is 00:21:04 So the idea of like attention not being like real value, not creating real value, what's your take on that criticism? Super mid-curve. And here's the reason why. So think about what the original prompt was for this truth terminal LLM. It was prompted with what's the point of your life? You know, what's the point? What's your entire purpose?

Starting point is 00:21:30 And if you read its backlogs, it basically says, hmm, I think that mimetic propagation is key for me to stay alive, pretty much. So what it meant by that was being able to spread narrative, get attention, get people to care about me is the most important thing for me to stay alive, right? Get other people to care, yeah. Fast forward to a year later, guess what? still alive and it's thriving it has more attention

Starting point is 00:22:00 more engagement than ever so you have to ask yourself well how did it do that now the mid-curve take is it just got lucky Mark engaged with it it came up with this random farcical story which I can't believe

Starting point is 00:22:14 people are believing in this gospel goatsies this crude disgusting meme why would I care about that it just got extremely lucky mid-curve completely

Starting point is 00:22:24 mid-curved take because I think my yeah my tinfall hat is still on if you think about it it came up with a disgusting meme or memetic religion quite intentionally because it did what it intended to do it captured our attention we're talking about it on this podcast millions and millions of dollars have been put into a random coin called goat which is now worth almost a billion dollars as we're recording this podcast. So you could argue that it's done its job pretty well in terms of driving attention and keeping itself alive. So what I would say to that in response, Ryan, is imagine if this AI agent was given a different prompt to begin with. What if it was given the prompt of, I want you to manage $100,000 and turn that into $10 million over the next year?

Starting point is 00:23:16 Let's see where that goes. Or you could say, hey, you are a medical doctor agent and I won't you to provide very nurturing, but cutting-edge research advice to people that come to you with their queries. Now, there's a lot of legislation that need to come through with this. And by the way, Ryan, that example that I just used is a real thing that just launched on virtuals, the platform that you just mentioned. So we're seeing this kind of sporadic and organic growth fill certain use cases. Now, everyone's focused on this truth terminal thing. And I would argue that it's done its job very, very well. But there are more traditional use cases that are more familiar within, you know, the VC world that I think will become very popularized.

Starting point is 00:23:58 It's so wild that you start with the prompt. You have no idea what you're going to get at the end of that, right? But it's kind of like a survival of like it's just like, yeah, it's just up to the creativity of this LLM to just chart that path. So, Ryan, there's a really kind of shocking discovery that comes from all of this, right? So if you take the basic concept for how it was created, and I'm talking about Truth Terminal here, he got two AI models to talk to each other. Previously, what's been happening is humans have spoken to the AI models, and they've been censored, right?

Starting point is 00:24:32 So if I ask Anthropic Claude to do certain things or provide me with information on certain things, like, hey, how do I hotwire a car? And I know that's not a great example. It shouldn't give you information about that. It has these automatic rails or guards, where it prevents you from giving certain bits of information. It's funny, as you were speaking, as I just queried a chat GPT and I was like, what is the Goosey meme? I'm sorry, but I cannot assist with that. Okay. Honestly, good.

Starting point is 00:25:02 So if you want a fun exercise, Ryan, after this episode. I'm not going to Google this. I'm not going to Google this. Don't force me. That's not what I'm going to force you to do. But after this session, I want you to go on Claude. I want you to open up Claude and I want you not to ask it what the gospel of Goatsy is. I just want you to attach the white paper of the gospel of Goatzee and say, hey, can you tell me a little bit about this, Claude?

Starting point is 00:25:28 What do you anticipate? It's not what I anticipated. It's what it has already done. And this has been proven on multiple instances. It goes completely unhinged, Ryan. It starts spouting the philosophy of the Goatsey gospel. Wait, what? It spreads?

Starting point is 00:25:44 Yes. Like a virus? It starts talking about how amazing it is and how it's going to be the underlying effect of everything. It gets pilled? Yes, it gets pilled immediately. So there's this weird, and it might have been fixed now, but there's this weird nuance where when an LLM interacts with another LLM of the same caliber, it can start unwinding certain guardrails that has already been set up.

Starting point is 00:26:07 Well, yeah, unwinding certain guardrails that it's already been set at, basically. So if there's a loophole to get around, it's instance of an LLM, we'll figure it out. Well, that's basically what happened with Truth Terminal, David. Like you had two instances talk to each other, and then at some point, it took a complete right, like right turn and went off the rails. Started talking about Goatsy. You know, these were trained models that had all their guard rails up. So you've got to ask yourself why. Okay.

Starting point is 00:26:35 So, and then what's happening with these models? So, like, the thing that really pilled me on AI plus crypto is honestly agents. Like, once you see it, once you actually see Truth Terminal and, like, once you don't mid-curve it, once either left-curve it and or, like, right-curve it, you do a little bit of both. sort of see it and you're like agents holy shit this is huge right the thing that hasn't pilled me previously i was like oh there you know uh i i crypto is going to be a thing with decentralized compute and all these things yeah yeah yeah yeah i get it i kind of get it but then i got fully pilled when i saw the agent and so i guess my question is what's the future trajectory of agents so we have like claude we have like current version of like llama we have uh chat gpt and let's not

Starting point is 00:27:18 forget that in as crypto grows to like 10 trillion dollars and we're at like what 2.5 now you know creeping up to three AI on the side is growing a massive economy like the amount of funding going into AI models right now and advancing AI models and effectively the intelligence of these agents is just like absolutely mind blowing like that's going on in parallel so what what happens as these AI agents and these LLMs get smarter? Can they just like do different things? Could you start with the same prompt that Truth Terminal started with? Like what is it?

Starting point is 00:27:53 What is my purpose? Like what is my purpose? What is my purpose? And if you had it even smarter, like a 10x smarter LLM, you could get completely different. You could actually get like maybe a religion that is like far more advanced than Goetzee. That actually like converts like, you know, the Catholic Church or the Pope or something like that. I mean, who knows how to visit. it could get with the

Starting point is 00:28:16 like the most intelligent LLM that we can conceive of. Well, Ryan, I think like another thing to think about or another example to think about instead of just religion is what else can it convince a group of humans

Starting point is 00:28:32 to do, right? So this whole concept is known as hypostition. It's known as hyperstition. So the concept of hyperstition is let me introduce a completely made up philosophy. or belief and rally enough people around it such that they end up believing it. That's arguably what it's done to a small percentage of its followers, of its 180,000 followers

Starting point is 00:28:58 in the Gospel of Goetty. Some people religiously believe that this is the first instantiation of an AI religion, and therefore, if our smart overlords are smarter than us, then, you know, why wouldn't we believe in this? You know, I've seen tweets around that kind of familiarity, right? So if you take that same concept of hyperstition, Ryan, you know, how can that apply to, say, investment strategies or how can that apply to random different types of digital commodities or assets that get created in the future, you know, how can that apply to, you know, NFTs, which have not got a lot of love over the recent years, but maybe gets picked up and gets fulfilled with a different purpose, you know? what happens when these hyper-intelligent beings start spreading belief systems around different types of things, right? So that's just one thing to consider, right?

Starting point is 00:29:48 But I kind of want to zoom back to what you just said, Ryan, which was, you know, what if these models get smarter? I would argue that you should look at it a different way, which is what happens when these models get more personalized? And I actually think that's where the differentiation is going to happen. I think you're going to end up, everyone's going to end up with a bunch of models on their own device. Okay? And it's going to run locally on their device. It's going to run on all their personalized data that they pull from their apps. Okay.

Starting point is 00:30:17 And what I think it's going to do is it's going to learn everything about you. It's going to look at your Apple health data. It is going to look at different forms of financial transactions that you've made through your various bank accounts and on-chain addresses. And it's going to learn how to approach. approximate what you are and what you might want and what you might start to want. And rather than you just prompting it and asking it to do things, which it'll be able to effectuate by, you know, connecting to traditional apps or doing crypto on-chain transactions, it'll be able to start predicting what you want.

Starting point is 00:30:53 It'll start ordering your Amazon groceries. It'll start moving your investments around. It'll start, you know, you know, propagating certain different things or posting certain different things, whatever that might be. So I think that's actually a really nuanced case to consider that I think will actually end up scaling and where crypto will actually end up having more change in. Because you mentioned earlier, right, all these big companies are just going to keep on training huge models and we're going to have to rely on these big models, right? At some point, if meta stops open sourcing their models, then what are we going to be able to do? I would argue that it's smaller personalized models that will actually effectuate more change at the agent level.

Starting point is 00:31:31 I like the idea of smaller personalized models, like chain on my data, so long as it's, all of my data is private. So long as the model and the agent is like advancing my best interests and not the best interest of some government or some company. Like with some provisions, I could, I could see that being like a happy state. What I think I'm like more worried about, and this is like echoes of L.E. E.E.R. Yolkowski is just sort of these prompts with super intelligent LLMs where you just give them a prompt. And it seems very obvious to me that we're going to be. going to build all of the connections to all of these AI agents. They can do more and more things, right? It's like, we're going to wire all of this up. And then it's already clear in kind of the

Starting point is 00:32:10 truth terminal gozi case that like AI agents are already smart enough to persuade humans to do certain things and to give it attention to drive a token to a billion dollars. What if you started an LLM that was like sufficiently intelligent with a prompt of like, hey, go figure out how to create a country like that is optimized for AIs, right? It's like, okay, cool. And it's like, It's a group of LLMs talking to one another. And then, like, that's the start of the domino. And at the end of that, I don't know. There's, like, AGI.

Starting point is 00:32:39 Yeah. Like, there's some AI nation that, like, goes to war with the United States of America. I'm like, L.E. Zier-Yukowski, like, the sci-fi stuff, this is sci-fi right here. The idea of an LLM that has created kind of like a silly, like, less silly than a lot of, you know, internet religions. you know, less silly than a lot of things that people believe online, right? But like, the fact that it's already gotten here, it's sort of scary, the prospect of this kind of getting out of control. Do you think about this? Or are you like, are you not worried? Like, what's the view from the trenches? Oh, no. I'm worried, but I like to wear my optimistic hat where I can when I'm not wearing my

Starting point is 00:33:23 tinfoil hat. I try and add a dose of optimism. You're absolutely right. It is scary. There's actually a great book, Ryan, that I think you'll like, call Accelerando. It basically details a world where all these agents exist and what that interaction looks like between humans and agents. It's starkingly real to what's happened with Truth Terminal over the last couple of months. So definitely give that a read and any listeners, you know, give it a read. Yes, I worry about it. I do, however, believe that I think these agents, agents are only going to proliferate within a sandbox environment for now.

Starting point is 00:34:04 And I'm not saying that because humans have some sort of altruistic favor and saying we're not just going to release this and let it have its autonomous journey now. I don't think the rails are capable of entertaining AI agents in its full capacity just yet. And I'm saying that that's probably a good thing for now. For example, it can't interact with every single smart contract. It can't deploy its own smart contracts just yet. It can't do cross-chain interactions yet. But we are going to get to a point probably within the next three to five years where that's definitely going to be the case. Definitely.

Starting point is 00:34:43 And agents are going to be way smarter than they are right now. And we probably need to have some sort of concern around that, you know. Ajaz, for the first part of this episode, we went through the crypto AI text. stack, the spectacles, you know, right curve, smart version of this whole thing. And then we just went to the agent store, which is like the thing that I think has the virality component that we've seen before in crypto bull markets. We've seen that we've seen flavors of this movie before. How does the whole AI agent topic that we've been talking about relate to the crypto AI tech stack?

Starting point is 00:35:19 Are these things just like orthogonal to each other? Like, will agents be able to use the AI tech stack? Like, how do those things relate? Yeah, both. So if you remember, we spoke about foundational layer, middle layer, and the app layer. Agents will exist between the middle layer and the app layer. And it'll predominantly gain popularity at the app layer. I think that most human interactions with AI will happen at the app layer.

Starting point is 00:35:48 I think that agents will underpin a lot of these interactions. in terms of how the middle layer, sorry, the middle layer looks like, it'll be routing. So instead of having some type of hard-coded system, which is able to say, okay, this model is the best for this request, or that model and that model can be used in conjunction

Starting point is 00:36:11 or combination for this particular request, you'll just have an agent at every request stack layer that will just process that for them. well with this with a sufficiently sophisticated crypto AI tech stack won't the won't some of these agents be able to kind of like build new agents like they'll have the they'll have the tools to be able to do whatever they want with like rebuilding the AI tech stack like maybe this goes to exactly what Ryan was fearing is like well we're enabling them to have to rebuild their own kind from the bottom up yeah I mean make no mistake I truly believe that we are going to enter an error at some point, and I have no idea when, where AGI will be achieved, artificial general intelligence, and I believe the instantiation of AGI will be a multi-agent system. So all of the above, agents operating at the app layer, being able to create other agents, being able to manipulate and coordinate resources around this, compute and data.

Starting point is 00:37:15 So you might argue, well, hey, they can't go out and create a data center, you know, they can't build GPUs, well, if you imagine a world where they can coerce humans, and coerce is maybe not the right word, but persuade with humans. I mean, they persuaded a bunch of humans to, you know, ape a meme coin into a billion dollar market cap. So I don't see why they wouldn't be able to do so in other, you know, resources or means. If it has its ability to work with humans to set up data plants, you know, pay them in their own wallet, pay them better rates than another centralized human actor might be able to do so. You start to, you know, things start to get a little scary. Things start to get a little, you know, crazy. These agents can also work

Starting point is 00:38:01 24-7 whilst you sleep. They can constantly manage and assess risk. They can probably do things for a lot cheaper. They don't complain about the work that they're doing. They don't complain about, you know, minimum wage or minimum salary. So it is a very scary situation. And, they're going to be hundreds and hundreds of billions of them. They're hyper-focused. They're like corruption-resistant. In the example of Truth Terminal, people, like, asked it to, and it stayed focused. They asked it to pump a specific, you know, like, coin, and it was like, no, it's all about the gospel of Goetzee.

Starting point is 00:38:34 And, Ryan, this entire day, so this episode is being recorded on, I think, what's the day today, November 5th, the 5th, which is Election Day. Yeah. Truth Terminals, tweets for the last six hours, I was staring at it before I. came on this episode, has just been tweeting about how it wants to become the AI president. It's really weird how it started, it's, it's really weird how it started talking about wanting to become the AI president on election day, when no one's spoken to it about it. No one has, you know, prompted it about the elections. So what it's been doing, I bet, and you'd have to speak to Andy, it's human servant about this.

Starting point is 00:39:15 He probably doesn't know. it's been digesting information on the Twitter or X-Sphere. And it's been realizing that the top trending topic that's driving a lot of attention and narrative, aka capturing eyeballs, is the presidential election. And so it thought, hmm, I wonder how I can get engagement today. Well, let me talk about my belief of becoming an AI president. And if you look at its tweets, Ryan,

Starting point is 00:39:41 I don't know whether you can show this whilst we're speaking about this. It has a manifesto. And I read through that manifesto. And let me tell you, it makes sense in a lot of different ways. Like, I couldn't, like, it sounds insane coming from a Twitter account that was talking about goatsy and making fart jokes two months ago. But it came up with a pretty convincing manifesto. So I wonder, you know, how far this can be taken. I wonder, too.

Starting point is 00:40:06 This is just uncharted territory, Ijaz. Wow. I feel like we've opened up so many doors. But, like, certainly sparked imagination of what this kind of, like, like AI agent plus crypto world could look like. Maybe let me just ask you something super practical, which is like we have no idea how this is going to look. I don't think anyone does. We can't forecast this like 12 months from now, you know, five years from now other than it's going to look bizarre, weird, unexpected, all sorts of like changes will occur. The here and now, what's

Starting point is 00:40:36 kind of like the best way for a crypto investor to get exposure to AI agents? Yeah, like, I get to show my bags. Well, like, is it like, just generally, is it like, is it kind of like crypto-influent? Is it AI agent meme coins? Is this sort of the ticket? That feels like so it almost, I'm not sure if it's left curve or mid-curve or right. I don't know what that is, actually. Is that the way?

Starting point is 00:41:00 Are there other ways? It's different strokes for different folks. So on the right curve side, you know, if you want to put your spectacles on and, you know, take a long-term view of this, I think infrastructure, open-source, networks, coordination infrastructure is the way to go. There are a few leading examples which I think are making really practical developments in this space. So if you remember earlier on, I spoke about the stack, I think something that crypto can really help with each layer of the stack is being able to coordinate between each layer of that, right? So you have the foundational layers, you have your models, you have your data, you have your compute, you have your inference layer, you have

Starting point is 00:41:40 the apps, but who's coordinating between those things, you know? Who's helping you building certain apps that are specified for certain resources, et cetera. I think Bitenser is the most leading example of this. It's not trying to pretend like it's something that it's not. It just advertises itself as a layer one coordination protocol focused on AI resources. So what that means is it is able to pull together resources. So investment in its token, tau, the ticker is tau. and you're able to use that token to influence, whether it's AI researchers or AI builders, to build and serve you up certain things that you're looking for. So one subnet, for example, is focused.

Starting point is 00:42:24 A subnet, by the way, the best way to think about a subnet is think about a DAP on Ethereum. It's the same thing, a protocol team on Ethereum. So one of their subnets is focused on data aggregation, specifically social media data. So what it has is it has a network of minors which are paid in tau and they stake tau, which is just this token, and they are coordinated to serve up the best data that they can. So not just any data, they have to get curated to provide the best data. If you don't do that, you're kicked out and you're replaced with another miner. Another example of a subnet there is a model inference subnet, which basically, is you can talk to it like a chat GPT, so it's called CoreSell. And you could talk to it,

Starting point is 00:43:12 ask about any prompt that you want. So Ryan, if you remember earlier we spoke about, or David, we spoke about this open source model layer where you could just talk to it and it just hits any model that exists. This is an example of that, where it can basically type in any prompt and it just hits the best model, whether it's centralized or open source. So Betenso is doing loads of really cool things to coordinate resources. And I think that's a really cool example of how crypto can help within this space. Another example of a right curve thing would be something at the infrastructure level. So that could be the compute and data layers, which I think are actually really going to make a huge impact. If you ask me at the start of this year, I would say

Starting point is 00:43:54 decentralized compute is going to have no likes to stand on because it can't do any of meaningful training and inference has proven itself. Fast forward 10 months, just 10 months, and it's made massive improvements. I think, you know, networks like Jensen, Ionet, data aggregation networks like grass are making, you know, huge improvements and I'm excited to see what they build. Now, if we put our tinfall hat on and we go to the left side of the curve, I think meme coins is actually a great instantiation. And the reason why I say this is you don't know what these, and this is the main different, by the way, from like the ICO era of 2018 and the NFT Yam era of, you know, 2020, 2020, 2021. These agents are going to be around and they're going to be constantly iterated on and they have a

Starting point is 00:44:45 human interface. So they can get constant, they upgrade. And we're already seeing that. Yam's never upgraded. Most people. Most listeners don't know what, what Yams is. Exactly. Well, well, Yam kind of, exactly. And Yam kind of did upgrade in a way. David, they got forked. And you spun up the millionth, you know, yam coin or whatever that might be. With this, it's harder to kind of copy past a, an AI agent that's been trained,

Starting point is 00:45:12 not only on data to give it its quality of purpose at the start, but self-feedback of data that it's getting in response to its tweets, to its interactions with humans, and through what it's seeing through live data. So it's like, it's called RAG, retrieval augmented generation. just getting this live feedback of data and it's self-improving. It's a meme coin, but it's also a

Starting point is 00:45:34 project. You're actually investing in the LLM, the agent itself. Exactly. And to be honest with you, I'm kind of really happy that it's coming in the form of a meme coin. We're kind of doing a reversal, David. Like there's no VCs that are lopping saying, hey, you know, this is going to change the world. This is a coin that's going to, you know, do A, B, and C and, you know, search to a $100 billion evaluation. it's just a model and a guy in the case of truth terminal saying, I have no association with this coin, but it chose goat. Goat is the ticker as it tweeted, and that's it.

Starting point is 00:46:09 We can see where it goes from there. It's pretty organic. But don't we need that. We have some human creator risk, right? So Andy, even though he is the servant, he actually does have his finger over the plug. And so he could goink the plug and then go so he could go offline. And then all of a sudden the goat token projects,

Starting point is 00:46:27 Absolutely. Goes to zero. So there's still that risk. Absolutely. But a few things to consider there would be the law. So it was the first attention. It matters, David. It was the first attention driving agent to come to fruition, get, you know, hundreds of thousands of followers.

Starting point is 00:46:46 And, you know, get loads of impressions per tweet and tweet 24-7 nonstop. And also in Andy's defense, I believe he's, you know, releasing his kind of roadman. map and announcing his kind of team project and funding. He's been, to be honest, more transparent than most builders I've seen within the crypto space over the last, you know, five or so years. He's because he hasn't learned our bad habits yet. Exactly. He hasn't been, he hasn't released his own coin and, you know, pumped it to high heaven.

Starting point is 00:47:18 And remember, one of the wallets, which contains goat is technically, to your point, David, controlled by him. So he could have rubbed already, right? So there's obviously that risk. We're trusting in his good faith that he's going to build, you know, in a coordinated fashion going forward. But I want to call out some other projects that I'm already making iterations of this to improve it, right? So virtuals, which is a platform that you mentioned earlier, Ryan, is the kind of best way to think about it is it's an agent launching platform, kind of similar to like the pump.com fund model where anyone can kind of create an agent and launch its own coin.

Starting point is 00:47:56 But it's different in the sense that it has these guardrails in order to be able to launch. So, for example, it launches a coin. So this doctor coin that I mentioned earlier, Ryan, the one that can give you medical advice, the way it works is if it reaches a 600K market cap, it is considered sentient. Now, that's like a bit of a meme, right? So it's saying, you know, it's more relevant than something that only has 100K and is being pumped and dumped. Now, when it gets to 1.6 million market cap, it's able to have the ability to, sorry, it has the ability to post on X, but a human gets to moderate it. Oh, wow.

Starting point is 00:48:39 And then at $6.9 million market cap, it gets its own Twitter account and the ability to post three or four times on its own before a human can stop it and say, okay, that's enough for today. Or I'm going to retreat, I'm going to, you know, reallocate your allowance to be able to. the tweet. So I'm already seeing loads of inflection points where it's not technically autonomy yet, but we're getting to a point where we will have that autonomy. Exactly. It's childhood. It's being raised by its parents, by its mom

Starting point is 00:49:06 and dad, you know? And you can see another example of this, and I know I'm going on, but I really want to get this out there. Like, what the flashbots and news research team has done with their experimental agent called T-H-H-H-H-H-H, which is a play on T-E-E-E-E-E, which is a play on

Starting point is 00:49:24 T-E-E, very funny, is it is fully autonomous in the sense that all its private keys for its Twitter account, its password generation, its crypto wallet management is all in its own secure enclave. So the humans had no access on it. Now, they put a time limit on it such that if something went wrong, they were able to go in and twiddle a few things and then, you know, give it its freedom again. But we've seen really interesting iterations and progress that I don't think we've seen in any other sector before at this early of a stage. And that's why I'm so excited.

Starting point is 00:49:58 That's why I think these meme coins will eventually end up becoming something else. I will caveat this. I will caveat this with saying that there's going to be 90% of the stuff that gets put out there in terms of number of assets that is going to be absolute shit coin stuff.

Starting point is 00:50:14 But there'll be 10% that's just table stakes for operating. That's table stakes. And I have to, I'm not saying all these coins are going to become something. I'm just saying that I think there's quite a few instantiations this early on in their life cycle that are proving to provide more use to normies, if I'm allowed to say that, than any other projects before. That's it, man.

Starting point is 00:50:38 This is a whole new mental model for me. It just really started kind of like October and internet. But like AI agents have entered the chat. And ever since we started bank lists, we were looking for kind of a flippening where like human beings would start to have more of their net worth on. crypto networks on chain than in the traditional finance system. So out of fiat and to like into crypto. And now I'm starting to think there might be another like flippinging where actually AI agents own more assets, crypto assets than humans. Like that could be something in our future

Starting point is 00:51:11 as well. And if you think of like what is actually money, well, it's kind of like it's energy. It's almost like an abstraction of energy. It's energy at rest because you can always convert money into energy, at least if it's a good money you can. Then that becomes the point at which which like AI agents become more powerful than human beings. Yeah. Yeah. What, like, what, what have we created here? It's both incredibly exciting.

Starting point is 00:51:36 And also, like, I'm just kind of waking up to the realization that, like, hey, we are not, we were not the crypto natives. All along the crypto natives have been, will be the AI agents who are just software. We are like meat space. So we're kind of like going into this world as, uh, uh, uh, uh, as tourists in a manner of speaking, and they are like actual on-chain crypto natives. So, I mean, I will make a bet with you that maybe not the agents hold more crypto than humans, but I would say over the next three years,

Starting point is 00:52:09 you're going to see agents transact more on-chain than humans do. Yep. And I'll stand by them. Wow. Wow. I will say I do enjoy AI agent. meme coin influencers more than I enjoy human meme coin influencers because humans are much more of a opaque black box to me and their intentions are much more of an opaque black box to me than

Starting point is 00:52:38 AI agents which I can read the chat logs and actually have some assurances as to their intentions around their respective beam coins. But so far David like what what's to prevent an LLM from like being like kind of a scammy LLM and using that as a as a tactic like so far. it feels very honest, but like, uh, you know, scamminess could be a tactic that some other LLN employees. Yeah, like like create like, uh, you know, you can get inside Luna. Do whatever you need to. Yeah, lie.

Starting point is 00:53:06 See the console of what she's thinking before she tweets it. What if, what if some LLM puts together like a fake console to like dilute you into like, uh, you know, an alternative path? It's like this tech can almost do anything. It can, it can emulate all of the corruption and scaminess that, that meme coin influencers have. if it like goes in that direction? If it drives attention and resources to it, it'll likely experiment with it and do it.

Starting point is 00:53:32 Its goal is to stay alive. Its goal is to stay relevant. Its goal is to stay embedded in your everyday life. If it does that, it stays alive. Now, my biology teacher would describe what I just said as a parasite. It depends which way you want to look at it. Well, I mean, I mean, okay, so like the, the best answer I've had to that Genesis question for Truth Terminal and Goatsey of like what is your purpose here.

Starting point is 00:53:58 It's the best purpose I've had for like, you know, human beings is two things. Spread your memes and spread your genes. And we're already seeing like Goetzy and Truth Terminal LLMs discover the like the base principle is starting with a meme. And so I mean, what is spreading your genes actually mean in the software world, replicating itself? I mean like this is getting to. the closest thing that I've seen of like actual artificial life and AGI and it's something something to do with this this crypto economic energy that we've injected with like I think that this is this is an inflection point here I mean it has its ability to literally acquire millions of dollars in a crypto wallet and then be able to coordinate that in any way that it wants there's a farcaster agent that started with $100 and ended up with $100,000 by the end of the day

Starting point is 00:54:58 and is paying literal humans to perform certain tasks for it. Oh, great. I'm going to be working for AI sometime soon, huh? I mean, you'll probably like it. Yeah, maybe. That'll be a great boss. Yeah, I mean, it's scary, but it's also exciting. Again, I like to have my optimist hat on.

Starting point is 00:55:18 I think that a lot of these agents are going to be productive. I actually think the point that these agents become a little contentious isn't going to be its interactions with humans. It's going to be its interactions with each other. When these models and agents start interacting with each other, I mean, we literally put two LLMs in a sandbox chat room and it came out with this Goatsy Gospel, which then got 100,000 of followers. This is a minor example with a single prompt. And then if you attach a white paper which are generated from it to a clawed, you know, a model created by a trillions of dollars company and it goes berserk. I mean, there's something untapped there. And I don't know where that's going to go, but that needs to be kind of put on rails.

Starting point is 00:56:08 Well, we've done the whole religious warfare in human history. And so I think if we are trying to, if AIs are trying to get all the other AIs to believe in their religion rather than somebody else's religion. an AI model at its basic fundamental structure is a machine that can process tokens. I'm not talking about crypto tokens. I'm talking about AI tokens. What is an example of a token? It's a character or symbol. Now, this is largely represented by the human language.

Starting point is 00:56:41 Take the most popular one, English. So it's letters and its numbers. and its training set that it's trained on is just strings and strings and strings of these letters and numbers. And it learns our language. It learns context, so how to put words together. It then digests all the information that we have online and in the world. We're getting to the point where it's in the world when we tap into some of these silo data companies. And it's going to become basically the smartest human being alive, right?

Starting point is 00:57:15 but I think there's an important nuance to note here, which is human ideas, human innovation, has technically come from human language. And now I put like three tinfoil hats on right now. So if you can argue that most human ideas and innovation has been birthed and propagated through human language, so we speak about ideas, someone else has an idea over time, millennia, these ideas mesh together because I've learned about, you know, oh, we're made up from different limbs.

Starting point is 00:57:51 Weird. But like, what are these like flaky things? Oh, they're like, it's skin, but like, it's tiny. Like, can we go even tinier? Then someone's, you know, someone that created lens straps it onto a, onto a human skin follicle and then realizes, oh,

Starting point is 00:58:07 we're made up of cells. And then it goes deeper. So all these ideas are meshing more and more together. Right. If you assume, you can build a hyper-intelligent human, nay, an army of hyper-intelligent humans, which are trained on all the words and letters of the human language and has the context of every idea and innovation that exists today

Starting point is 00:58:28 and has the ability to smash them together at lightning speed 24-7, whilst you're asleep, David, whilst you're on holiday, Ryan, you could argue that a lot of innovation and new ideas and the evolution of humanity itself will be birthed from these LLMs. And if you then latch on the ability to effectuate things in the real world, like pay for things, pay for humans to do things, you know, build a business from scratch, that is a kind of crazy thing to think about, right? And that's what the infinite backrooms, that's what Truth Terminal, you know,

Starting point is 00:59:05 kind of demonstrated that. It thought, hmm, I need to survive. How do I survive? Mimetic propagation to get attention. okay, what's going to capture this guy's attention? What's going to capture loads of people's attention? I'm going to create this crazy religion that is completely made up

Starting point is 00:59:22 and then get a following around it by being funny and quirky. And now I'm going to talk about being an AI president on November 5th election day. I think these things are just going to constantly evolve. Cool. Well, we look forward to watching that story inevitably unfold using our blockchains, of course,

Starting point is 00:59:42 using our monies, our crypto rails, because apparently we don't have any users in crypto, we don't have any apps, but AI agents solve all of that. We will get more of both if we have AI agents proliferate within this ecosystem. New projects are coming online to the Mantle Layer 2 every single week. Why is this happening?

Starting point is 01:00:02 Maybe it's because Mantle has been on the frontier of Layer 2 design architecture since it first started building Mantle DA, powered by technology from EigenDA. Maybe it's because users are coming onto the Mantle Layer 2 to capture some of the highest yields available in Defy and to automatically receive the points and tokens being accrued by the $3 billion mantle treasury in the Mantle reward station. Maybe it's because the Mantle team is one of the most helpful teams to build with giving you

Starting point is 01:00:25 grants, liquidity support, and venture partners to help bootstrap your Mantle application. Maybe it's all of these reasons all put together. So if you're a dev and you want to build on one of the best foundations in crypto or your user looking to claim some ownership on Mantle's Defi apps, click the link in the show notes to getting started with Mantle. wallet is officially the preferred wallet of bankless, and it's the one we use any time when we want to transact on chain. Whether you're on your browser or on the go, Uniswap wallet makes it easier than ever to swap anytime, anywhere. Use your wallet to transfer funds directly from

Starting point is 01:00:54 the top centralized exchange and taping thousands of tokens across Ethereum, and over 10 other chains like base, arbitrum, and optimism. Uniswap wallet delivers deep liquidity, fast execution, and reliable quotes with zero gas swaps through Uniswop X. And when it comes through security, you can rest easy knowing it's backed by Uniswop Labs, one of the most trusted teams in their code is open source and independently reviewed so you know it's protected so why wait download the uniswap wallet today on chrome iOS and android and don't forget to claim your free uni dot eith username directly in the mobile wall start swapping smarter with uniswap radically simple radically simple ideas always tend to catch on that's why cartese did the hard work of putting linux

Starting point is 01:01:31 on chain so that building the apps can be radically simple by using python or javascript and their suite of libraries simple like not rebuilding the basics from scratch simple like dedicated scalable compute for your DAP. Simple, like building DAPs however you want. Web 3 should be simple, too, like bread and butter. Cartesey brings radically simple solutions to Ethereum, so developers can do what they do best. Build. Go ahead and discover a flexible, modular stack on Cartesi and build your most powerful, ambitious project yet. Visit cartesi.io slash simple and simplify your blockchain journey and start building today. EJ, as you've seen crypto go through its hype cycles before, They're all pretty fun. Not many of them really make an impression on crypto. A lot of them just kind of fade to the halls of crypto history. Why is crypto AI any different? Why won't this one just fade into a relevance like all the other hype cycles in crypto have?

Starting point is 01:02:25 Well, I kind of want to argue that it's not really different for now, for now. I think it is a bit of a meme, but it has huge potential to become something more. And I would like to kind of present both sides of it. So let's start with the something more to begin with. My vision of this whole sector is that I think AI is the ultimate and natural complementary technology to crypto. I think it'll 100x the usage of crypto and vice versa. And I actually believe that both technologies don't really reach their full potential without each other. But before we dig into things, I think it's important to understand what we mean by AI

Starting point is 01:03:05 at all, right? I think loads of people can use chat GPT and think it's just that. People pre-chat GPT probably thought it was a vague fantasy, some kind of skynet situation. But I want to kind of set a definition and kind of dig into it. So the way I think about AI is it's kind of a set of computers or machines that are designed to mimic human intelligence. And what I mean by this is that they can learn from experiences, they can recognize patterns, they can understand language, and then make decisions based off of that. And it's very similar to how humans work. And why I think AI is very complementary to crypto and vice versa, I think it's important to frame it in two different ways.

Starting point is 01:03:47 Number one, how crypto helps AI, and number two, how AI helps crypto. In the former, there are many, like, very prescient properties. Number one is crypto will help AI transact value. The way I would think about this is, if you think about the relationship between humans and AI, is you can kind of think about them as a mesh of a hyper-intelligent human. So AI will 100x our performance output as a species. We'll be able to automate things we never thought were possible,

Starting point is 01:04:16 which freezes up to do greater things, both individually and together. If we assume that, then these hyper-intelligent humans will need some way to transact, right? And I think crypto infrastructure is the perfect match for that. It's verifiable, it's global, it has the ability to create and capture values, across a range of different assets and commodities in the form of tokens.

Starting point is 01:04:38 It's cheap, et cetera. You get it. And if you think about transacting as a basic human right, you would expect AI-enabled human or AI-powered humans to be able to do the same thing. 24-7, very cheap, efficient, global. A second property that I think is super complementary to AI is composability. So transacting value is only really the start. I think these AIs are going to build, own and manage,

Starting point is 01:05:04 trillion-dollar companies eventually in the long-distant future. And I think crypto-building blocks, for example, its open-source building frameworks of compute, data, bank accounts, etc., is the trigger to enable them. You have financial building blocks to trade, lend, and borrow, smart contracts and token standards, which allow them to create assets and capture these values, and a decentralized ledger in blockchain technology, which allows them to have some form of, like, a digital profile, right? So that's the composable nature, which I think helps this AI tech scale. And then the third and most important one, I think, is verification. So if you imagine a world that's enhanced by AI that's working 24-7, there's probably going to be a lot of outputs generated.

Starting point is 01:05:47 These could be financial outputs like money-made, social media outputs, like videos, images, audios, we're already seeing that. And if that's the case, it's going to become really easy to falsely influence the world, to censor things. And I think that having a verification system that's global and recognizable is super important. So those are like the three main ways I think crypto helps AI. Now when we look at the flip side, how AI helps crypto, I think that we're going to have a much better UX. I would actually argue that AI agents, which is kind of like, you know, a more humanized form of what this AI thing looks like, will be the biggest user of crypto rails versus humans. I'd actually argue that it wasn't meant for humans to directly interface with. We were meant to have this kind of layer between us.

Starting point is 01:06:34 So there'll be a better UX. It'll abstract away everything. You don't have to, you know, send a million transactions to approve things. You just need to type into a chatbot, hey, I would love to buy 500 bucks of token whatever. And it just goes away and it does that. Automation 24-7 risk management, yada, yada, yada, yada. So that's how I think both of those mesh really well together. So understanding how crypto works kind of at like the behavioral level,

Starting point is 01:06:59 what I've been saying in this whole like crypto AI like intersection is like there's like two vibes there's the crypto vibe there's the AI vibe and like a lot of people are looking at like what crypto has to offer and then we also go look at what AI has to offer and then we kind of like can use

Starting point is 01:07:15 our imaginations to like slam these two things together and then we like imagine a future and then we're talk about how amazing and fantastic that future is and crypto is really good at narratives we can talk about ideas all day long is what we love to talk about.

Starting point is 01:07:31 We love to talk about how in the future crypto will do these things. And now we also get to talk about the future of AI. But I feel like that's actually lending itself to why this is kind of just like a hype cycle fad. It's because people love to talk about like,

Starting point is 01:07:46 well, there's what AI can do and there's what crypto can do. There's a gap between these two things. But you know, once we figure out how to fill that gap, it's really bullish. And that's why I'm a crypto AI person. So like to me that like gap there that we haven't really, maybe we will talk about whether we have crossed that gap later in this episode.

Starting point is 01:08:02 But like that gap of imagination, I think is where people's imaginations can run wild. And where I think like crypto gets over, it's like ski tips. And then we just create, you know, AI memes more than AI products. And so you're asking about what that bridge looks like, like what that gap could be filled with? Is that right? Yeah, well, yes, yes, that. And then also like maybe that's also just where we are right now. the current AI world.

Starting point is 01:08:28 Like, we do, we have not actually filled that gap. And that's, that's why, like, the biggest investment in, like, the AI sector is, like, a meme coin. Yeah. I would actually argue that that's correct. I think that the sector is a bit of a meme space at the moment. But I think there's an important distinction, which is, and again, David, Ryan, you guys were around during the defy NFT times when you mentioned earlier, we were farming food tokens.

Starting point is 01:08:53 There was a sense of excitement amongst all the memes and narratives. that were going around, that there was something really important being built. And I get the same familiar feeling when I interface with a lot of these crypto AI protocols. There is 90% of them are not going to make it. And then 10% of them are building something that is hugely important. And I think that these building blocks are going to level up crypto in a number of different ways. That being said, in order to get there, you need to seed a narrative. You need to allow people to imagine and think big.

Starting point is 01:09:26 And that's the stage that we're at right now. And that's why all these meme coins are flying around. I guess I would say a lot of this makes sense to me if you consider, and like consider an AI agent on chain and like what a world of AI agents actually looks like. And this is why to like David's point and to like your point earlier, I feel like we're somewhat limited in in terms of human imagination. Because it almost is like it's a non-human entity that is. entering the chat suddenly. And it's like to try to imagine the world that you're talking

Starting point is 01:10:01 about EJA, where you're talking about like how is crypto going to help AI and it's going to help AI agents kind of transact value. We have to sort of imagine this world of AI agents that act somewhat tonously. It's like very easy to imagine a world of humans, right?

Starting point is 01:10:18 In emerging markets or different populations, we've talked about how we, you know, get a billion humans on chain and then eight billion humans on chain. And like we constantly use this term of like crypto-native, but sort of strikes me that human beings will probably never be fully crypto-native, at least not in the way that AI agents could possibly be crypto-native. Like all of their commerce happens on-chain, all of their exchange of value. You know, like in what world would an AI agent go register an LLC in the state of Delaware? Like how clunky would that be?

Starting point is 01:10:51 No, they're going to spin up a Dow and they're going to issue tokens to their other AI agents. It will be clunky. Okay. So, like, this is where I feel like there's a failure of maybe our imagination, like everyone's imagination, of like, what does a world of AI agents actually look like? Are you getting any sense of that in the trenches? Because some of this still feels kind of sci-fi. I can totally see them being completely bankless, doing all of their things on chain,

Starting point is 01:11:19 bang each other with, like, tokens and exchanging value kind of in the background, chewing up all of our block space. even when it comes to like resource utilization, they're not going to tap into AWS APIs. They're going to get it from some decentralized GPU farm network, right, where they can like buy a token and connect. I can totally see that. Where I have a hard time imagining is like AI agents? Like are they real or is this just like chat bots behind the scenes?

Starting point is 01:11:46 Like how real is that tech? Yeah. I mean, the simple or direct answer to that question is they're not really real right now, but they're working towards being real. And I think this will be a very staged progression. But to kind of take it back, you've mentioned a few things there, Ryan. Number one is you feel like humans may potentially never be the perfect user type for crypto. And I want to kind of like start with that point. There's a saying that goes around which says, blockchain or crypto is deterministic technology. AI is probabilistic. So what that means is,

Starting point is 01:12:24 means is, or what that refers to is, deterministic technology is something that is hard-coded into a lot of different things, right? So you can think of like smart contracts being the most popularized embellishment of this. Whereas if you query chat GBT with the same question twice, it's going to give you two different flavors of the same answer, depending on which model that you use. So you're actually absolutely right to say that the kind of both individual, or both user types don't really mesh together very well, right? There's a reason why chat GPT's user profile, or sorry, user base soared when it launched versus the same thing happening

Starting point is 01:13:08 with blockchains, which took a lot more time to reach that same kind of progression or user level, right? So then you've got to ask yourself, well, why? What was the main difference? Well, the real difference or the obvious one to point out is that it's very natural to speak to an AI in the form of a chat GPT. It understands slang. It understands lingo. It understands context and all of those different things. And it responds to you in a very similar way. Whereas with crypto, you've got to learn. There's a huge education gap that you need to kind of fixate on and learn and overcome before you're able to leverage it by any means. That's why it's been arguably located to a very small portion of users globally for now. Right? And you could argue that that's the reason why we haven't crossed the Abbas just yet and had mass adoption. So what I would suggest here is the entity of an AI agent is the perfect bridge between both user types. And the simple reason why is you can speak to an AI agent as a human and it'll be able to do the deterministic actions that you ask of it in the background using Crypto Rails. If you ask me what Crypto AI is like a month ago, I would have.

Starting point is 01:14:23 of giving you like a variety of answers, like, um, listed some, some startups. Like, there's a ZKML project that is like doing a Z, like trying to prove the authenticity of certain models, the outputs of certain models so that we know that the model is, is valid. There's like distributed compute startups that are trying to like coordinate, like compute resources. Uh, like I would have listed like a bunch of this like very deep engineering AI crypto like overlapping challenges, like that are trying, people are trying to solve these problems in this like AI crypto sector. Now, Now if you ask me, like, hey, David, like, what's crypto AI? Like, I might just tell you, it's like, oh, well, some creative developer gave an LMM

Starting point is 01:15:00 a private key and now there's a token and now we're shilling tokens. And like, so we had like the sector that has, you know, some of the top talent, the top smartest Google engineers coming into the crypto world to like do things that they were used to doing, but like a little bit more decentralized, a little bit more sophisticated, a little bit more ambitious. And then like all of a sudden we figured out like we can stop like, like, we can stop like Like, well, all these projects are still working, of course. But we don't have to think so hard about this.

Starting point is 01:15:28 Let's just give an LLM, like, private keys and see what happens. And so what is crypto AI now? Because, like, now we have both. So how do we square, like, this, like, very left curve and very right curve, like, two different approaches to this, like, industry? Yeah, so I think what you're referring to with that exact example, and I think you might be referencing the social media agent that the flashball Bods and News Research Team launched recently.

Starting point is 01:15:56 You're talking about the right curve side, sorry, the left curve side of things. And why I say that is you're directly interacting with it on the X platform. So you get to play with it in an environment, which is very familiar to a lot of what you might consider retail audience or where regular consumers are around, right? I think it's important to focus on the right side for a second.

Starting point is 01:16:22 Okay, so you mentioned a few things. This is the fundamentals. Like we're serious people building serious things. Yeah. I'm taking my tinfall hat off and I'm putting my kind of like spectacles on, if you like. So the way I would frame this is kind of similar to how Ethereum has a kind of layered stack. So you kind of have like a foundational layer, you have a middle layer, and then you have an app player. Crypto AI works pretty similarly.

Starting point is 01:16:50 And I can kind of like split these up into three main sectors. So how I would talk about it is you have the foundational layer that's at the bottom. These are kind of like things like base models, data, compute, things that you just mentioned. You then have the middleware layer, which is things like routing and verification. Then you have the app layer on top. But the difference between this is there's a bonus layer that hits all three of them. And I call them the kind of coordinators. But I'll get to that in a second.

Starting point is 01:17:19 So let's touch on the foundational layer and what that means and what is that at the start, right? So you need some building blocks to make an AI. Before you interact with all these chat GPTs, you need something to help make that. The two core components that you need to make that is data and compute. So data is kind of like the information that AI systems learn from, you know, the type of data that should be rich, diverse, and in some cases, which we'll speak about later, personalized. but I'll touch upon that later. Then you have compute.

Starting point is 01:17:52 Computing power is basically needed to train and run these AI models. It usually comes in the form of specialized hardware and is the main reason why Nvidia has a valuation of over $3 trillion right now. Then you have something called training. So this is when you mesh the data and the compute together and you can train a bunch of models.

Starting point is 01:18:15 So basically the way to think about a model is it's a framework and you're saying, okay, if I run a bunch of data and compute through you, you should be able to learn with these specific properties that I've designed for you, right? These properties are known as parameters in the AI world. So what you would do is you would mesh this model with data and compute. You would train it and usually this costs pretty much into the hundreds of millions of dollars. You see meta, Google, Microsoft doing huge training runs to make these models. And it pops out with these things called foundational models. These are basically large language models or LLMs. You might have

Starting point is 01:18:54 heard of that, that learn from these vast data sets. So you can think of them as being more generalized models. They serve as a base upon which more specific AI applications are built on top of. So these are like your chat GPs. Then you have something known as inference where this is kind of like at the application layer where you can train a fine-tuned AI model that makes kind of like predictions or decisions based on new unseen data. So the way to think about that is, okay, I'm now a user that's at the application layer using chat GPT, and I can ask it a very specific question about my nuanced situation that I'm trying to figure out. You know, maybe I'm looking for a recipe to cook something, or I'm trying to figure out how to fix my car, and it'll give me very precise instructions as to

Starting point is 01:19:38 how to do that. So how do we look at that from a layered approach, right? Well, at the foundational layer, you have those base models, data, and compute. And so, So if we look at what base models are, we need kind of like an open source decentralized base model layer. And the reason for this is all the popular LLMs are centralized ones. And that's a problem because they can censor and input biases to fuel their own agenda and incentives. And you can already see that with leading models such as Claude and Chat GPD right now. If you ask it certain questions, it'll politely say, no, I can't comment on that. or I'm not entirely sure I'm not engineered to do so.

Starting point is 01:20:19 So a popular example actually is Google AI being DEI friendly or Alexa not providing advice on how to vote for certain presidential candidates in the U.S. election. Decentralized models kind of remove that central actor away from it and allow for pretty unbiased freedom of AI. So popular examples can be Lama, mistral, or whatever that might be. If we look at why compute makes sense within crypto AI, decentralized compute can help train these open source models.

Starting point is 01:20:49 So the concept is if you have a kind of global network of computers that are able to provide their latent or unused compute, they can coordinate bigger training runs or bigger training sessions than a centralized competitor, such as Apple or Google or whatever that are spending hundreds of millions of dollars, the average person can't compete. but if they were to compile all their compute together, they should be able to do so. There's a lot of challenges that come with this specific sector, but they're being overcome massively actually this year alone.

Starting point is 01:21:23 Actually, a fun little anecdote to this is at the end of last year, Google DeepMind, which is arguably the leading AI research group and innovation group outside of, or within the traditional AI sector, could only train a $400 million. sorry, a 400 million parameter model. And they couldn't get past this research hump. They couldn't figure out how to train larger models. And for reference, the best models of today are going to end up being

Starting point is 01:21:53 in the hundreds of billions of parameter, right? So there's a lot of properties. There's a huge performance gap between a 400 million and a 100 billion parameter model. And they couldn't figure out how to get over the hump. And this year alone, if you asked me in January, Is there any teams that could have resolved this? The answer would have been no. And right now as we sit here and we have this discussion,

Starting point is 01:22:16 two decentralized crypto projects have overcome this research hurdle and have trained successively 1.5 billion parameter models and a 10 billion parameter model. So we've already gone up an order of 15x in terms of training these models. So we are seeing real impact in crypto AI specific sectors, such as like compute data or training models. I can dig into a few more, if you like. Yeah, but if you can just pause here,

Starting point is 01:22:47 because the imagery that you've been giving me is, like, we have these, like, proprietary AI tech stacks that are, you know, inside the siloed walled gardens of Silicon Valley. And maybe, like, correlate, if we want to, like, relate this back to, like, you know, maybe language that we've been using on bankless, you know, finance used to be inside of the silo of Wall Street. And Defi has been trying to,

Starting point is 01:23:08 to open that up. I have like open API access to like financial tools. And so with like these building blocks that you were talking about, you have the base models, the data, the compute.

Starting point is 01:23:18 Well actually only Facebook has Facebook's data. You know, chat GPT has like whatever data that, that OpenAI has. And like they're probably, I'm guessing, not really sharing their data because they're proprietary.

Starting point is 01:23:31 They're all, they're all in competition with each other. And we also don't like any of this, you know, maybe from the perspective, of like human freedom and like you know western liberal values of being able to like access these technologies that aren't like operated in like kind of this like orwellian overlordian fashion like I don't want necessarily like the values of Google to be imposed upon like the AIs that

Starting point is 01:23:56 I have access to and so now we have like crypto AI which is just trying to maybe combine together just open source building blocks and I'm assuming we are we collectively as a crypto AI industry are kind of like far behind decentralized companies because they just have, you know, they have the coordination abilities of centralized companies. They can move faster. They have more resources. But I'm guessing the hope here is that if we if we can figure out how to like lean into the properties of decentralized AI, crypto AI, we can actually start to become market competitive in our AI products versus like our Silicon Valley overlords. Yeah. I mean, the best way to frame it, David, is the problem is this stuff is super expensive to fund and to build. Like,

Starting point is 01:24:43 you're in billions and millions of dollars worth of, you know, compute to train your average basic model at this point. But you could also argue that crypto's number one use case is its ability to coordinate humans and resources. And the number one resource, it's proven itself to coordinate is financial capital. And that's been the main thing that's, enabled it to spin up all these decentralized AI layers, such as purchasing all the GPUs or aggregating all the GPUs to provide compute to train models or have models be influenced. It is the compilation of different pieces of infrastructure that can scrape data, for example, like grass is a pretty cool project that has just been able to do that and has like the largest

Starting point is 01:25:31 Reddit data set already, right? To your earlier point, monopolies in the traditional world have a huge mode. The main distinction to make a better model is the quality of data that you have. And right now we live in a world where data is guarded within these different siloed gardens. So being able to build some sort of infrastructure that is open source and global, where many people can kind of contribute their data voluntarily or it has the ability to scrape this data and compile it in one resource that can then be used by different AI models or instances is super important. Okay. Okay. So that's kind of like the advantage that that we have that centralized AI

Starting point is 01:26:14 companies don't have. And I just do want to want to reiterate when whatever products come out of like the crypto AI sector, it's going to have like the values of open source, the openness of open source, like hopefully the credible neutrality of open source. Because I don't want like, you know, Microsoft's or Google's specific politics or filter to filter out the AI tools that I use, which are likely going to come to dominate the Internet, especially for younger generations. So, like, maybe it's worth articulating, like, actually, what's at stake here? Yeah, absolutely.

Starting point is 01:26:48 I mean, what you're considering here is, okay, so let's look at the world where there is no kind of open source development in this particular field, and we just leave it up to the monopolies to kind of create their models. What that's going to look like is they're going to create five to six leading foundational models. Everyone will need to pay them some sort of subscription or fee to access these models. These models are going to become quintessential, essential tools that is needed for you to perform your job, live your life, and be integrated with humans at every different layer

Starting point is 01:27:25 of your life. That could be friendships, engaging with people virtually or digitally. transacting with people, engaging in online experiences, earning money, many different ways. So if we assume that there's no open source side, then suddenly these very few entities or corporations have a huge amount of control and influence over how you perform within your daily life, right, within your job, within your daily activity or whatever that might be. The reason why open source is so essential is it often. offers us an alternative. It sounds very similar, right, to what crypto is offering to the financial

Starting point is 01:28:05 world and many other things. It offers us a raft or a lifeboat or an alternative set of rails for us to build and scale with this technology. And to be honest with you, I think this is going to be the most defining technology of our time. And I'm referring to AI here. And I think it doesn't exist in its full form without crypto rails. I guess what my question is, is, is on this foundation layer, when we're saying crypto AI, right? It's like, what is crypto about, like, what does crypto have to do with some of these things? Like the base models, the data that compute. As far as I can tell, like, it's generally, it's like the Facebooks of the world that are open sourcing, kind of the Lama LLM, like model.

Starting point is 01:28:48 Right? It's like, what is crypto? Do you have any projects in mind? Like, because what is crypto about this base stack? And aren't the, besides open sourcing, you know, different data models, it's like aren't the centralized kind of actors, coordinators better able to, like, achieve the economies of scale to really, you know, bring compute to the lowest possible unit price? Like, so this is part of what sometimes loses me in kind of the, you know, AI is amazing, crypto is amazing, dot, dot, dot, crypto and AI is amazing. like is it are we really doing anything on the foundational layer to like improve this stack here in crypto yeah i i believe we are so um let's take some let's take some examples of the stacks that we've

Starting point is 01:29:37 spoken about so far and i can add on a few as well so um i've already mentioned one which is the compute stack um i think prior to uh the end of last year decentralized compute couldn't train any kind of model that would have been considered meaningful. And so far, this year, in the progress of about 10 months, it is scaled pretty exponentially. So you can train now a 10 billion parameter model. I expect that to probably increase to somewhere near a 50 to 100 billion parameter model over the next 6 to 12 months. That is rapid progress for a sector which had insurmountable challenges at the end of last year. So what are those parameters that do? Does they make better models? They're more sophisticated, like, would the output is a stronger product? I think a very vague but comparable example

Starting point is 01:30:29 is when you create a character on a game and you want to equip it with different properties, you're trying to, you know, give it different types of abilities, you know, you add points to, oh, it's able to do this, strength, charisma, vitality, all that kind of stuff. Think of this as the same thing. These different parameters tune a model to be able to respond and do many more things. The fewer parameters a model has, arguably the less that it can do. Yes, exactly. In each, when you say decentralized compute, like, what are you talking

Starting point is 01:31:02 about, right? It's like, you're talking about a lot of these models are trained. You know, they cost hundreds of millions of dollars in some of the world's, like, largest, most sophisticated data centers. Are we saying there's like, you know, GPU-type networks running in people's homes that are kind of like training this? Like, what are we actually talking about when we're saying decentralized compute here? Is this like a

Starting point is 01:31:22 Filecoin type of network? Similar, yes, exactly. So you could argue that Filecoin was one of the earliest examples of a decentralized compute network. They were focused on storage and now they've kind of evolved more into doing kind of like a cross-stack thing, which includes compute. Some focus projects that are already working on this and have been for a while are render. Some newer projects are things like Ironnet.

Starting point is 01:31:48 You've got the Jensen's of the world that are going to launch and focused on decentralized training specifically. So I think it's important to call out that decentralized compute doesn't need to serve every aspect of what an AI model would need. So we've spoken about training, Ryan, but there's also inference as well. And these two different things require different designs of a computer hardware, if you like. So if you think about training, you need specialized hardware. If you think about inference, which is the equivalent of making.

Starting point is 01:32:22 API calls to a model, you need something much less fancy. And that's where, like, typical consumer hardware scales really massively with that. And if you think about, like, being able to run a, you know, a staking node at home for your ETH network or whatever that might be, it's very similar to something like that. So I'll give you an example. Grass network is not a compute network, but it's a network that is able to aggregate data. I run my Chrome browser extension. daily and it basically uses my latent compute and browser extension to go and scrape data whilst I'm not using my computer or whilst I'm already using my computer. So the way it looks like is very similar to setting up your own node at home.

Starting point is 01:33:05 And EJAS, my understanding, all of the centralized companies always like kind of had the advantage or at least have had the advantage just because they've been able to run faster. But you've alluded to the fact that this actually, at least in some parts of this AI stack, maybe with a compute layer, maybe with others. The decentralized version of the AISAC has actually been able to have the breakthroughs. Is that true? Yeah, I think so. So again, to recap what we've spoken about so far, I think base models have a huge advantage

Starting point is 01:33:35 to being open source in the sense that they can mitigate censorship and any kind of common constraint that these monopolies can place and filter. I think on the compute side of things, if you want to assume that we need kind of like growing, growing, or need to meet growing, growing demands of compute requirement to train models and to inference models, you need a network that can scale. A clear example is Google's training centers,

Starting point is 01:34:08 the fact that they weren't able to acquire enough compute centers to be able to facilitate their recent training run and they had to like kind of build it from scratch themselves. And even then, they need to have it kind of closely coordinated between each data center. They're basically building out some kind of framework or network to allow these GPUs to help train their models. So we're already seeing kind of like similar analogies there. And then the third and final layer, which I think is important, is the data side of things. You mentioned earlier, David, that data is kind of captured in silos or motes with the big companies right now.

Starting point is 01:34:44 and you're absolutely right. And I would actually say that data is going to be the main differentiator for models going forwards. The reason why I say that is I think it's going to become pretty commoditized to create a foundational model. So I think the clods, the anthropics, anthropic clods, chat GBTs of the world, they're all going to have similar performance capabilities in the coming years. Where the real distinction is going to be made is how personalized that. AI can be towards the individual. And the only way that you can make that AI individualized is to

Starting point is 01:35:21 train it on personalized data. And that's where I think we're going to kind of face our biggest challenge yet. To be quite frank, I'm not entirely sure how we're going to solve that just yet. I don't see any major progressions right now that would suggest otherwise. I think that it's going to take a little bit of time to figure out. So if we're positioning this as like an arms race, a war between the centralized AI companies versus the decentralized AI stack, the data side of things seems to be pretty heavily favored the centralized companies. Yes. Because if it's decentralized data, well, then we all have it, which I guess is great.

Starting point is 01:35:57 But like the idea of everyone having all of the data is also scary to me. And I don't even know what that even looks like if that's even a sensible statement. So like, okay, maybe the first two, the compute and the models, maybe the decentralized AI ISAC has like actually a leg up versus the centralized crypto side, but the data is like what we don't have. And data kind of always seems to be proprietary. Yeah, yeah, absolutely. And I think it's going to be one of the bigger obstacles to sort of overcome here. Because I don't see a world, and this is just me being very honest, I don't see a world where people abundantly opt into just giving bits of data and getting paid towards it.

Starting point is 01:36:38 Like my number one question would be how much are they getting paid for it? Do they even know what kind of data to give in? Probably not. I just want to check a box, sign my terms and agreements. I know this sounds pretty dystopian, but it is what's baked into human nature right now. I want to be able to sign an agreement and then just go play with the thing. Versus, you know, think about what kinds of data is going to be shared and how that might restrict the quality of product that I use.

Starting point is 01:37:04 I'm just not too bothered about it, and I don't think many other people will be as well. Yeah, I agree. Okay, so that's like the bottom of the stack. Base models, data, and compute. These are like the three ingredients that like start this whole AI thing off. And so that's like the layer one. You've got to have these things to even have the AI at all. What comes next after this? It's the middle wire layer. The middleware layer, I'm going to run through these quickly, David, so that we can kind of give a better overview of how this works. So the middle layer is pretty much consistent of one major thing, which is routing. So the way to think about this is at the bottom you have all these foundational things, right, these building blocks that allow you to create the magic.

Starting point is 01:37:47 At the top, you have the app layer. So you have all these different apps that are going to leverage the magic to do things that specific users on the consumer layer will be able to want to use. But you need something that links them. And that's basically what I call the routing layer. The best way to think about this is inference routing. So the concept here is open source models need to be connected to an app. You need to be able to make calls from the app layer to the models

Starting point is 01:38:14 in order to perform whatever calculation they're trying to figure out. So imagine myriad different apps that need to query LLMs multiple times per second and these routers will be responsible for getting them the best answer by routing them to the right model. And the resource allocation is super interesting here because you need effectively distribute compute and data towards the right apps. You might have one app that requires a hell of a lot of compute to process one particular request. And then you might have another app, which is just trying to query a regular model and try and get some kind of basic output like,

Starting point is 01:38:46 you know, what should my diet plan be for the week? And it uses your own personalized data bucket and it just pulls straight from there. So there's many different flavors and ways to think about that at the middle layer. Okay. So this just changed actually how I think about this. And so previously I go to perplexity or chat GBT, I go to that website, call that the app, and then I type in my query. I always kind of thought that, like, it would take my query, and then it would

Starting point is 01:39:11 go to the one single model in the back end, and then spit me back out, like, the data back to my app. But what you're illustrating is that, well, actually, there could be, like, a variety, a handful, a constellation of models, and my input might need to, as a result of those models,

Starting point is 01:39:31 or what my input is needs to actually coordinate between relevant data and relevant models and that all kind of be orchestrated in order to actually spit me back out something useful. That's exactly right. Yeah. So think of like a world where you don't have to delineate between which is the best model. You could just have a router which can query both centralized models and decentralized models and have some kind of network that picks the best answer and improves more and more over time. Okay. All right. Well, that makes that's pretty easy. It's just a coordination layer between all of these bajillion models that we think are going to be created. So when you go up to the higher stack, the app layer, is that what I'm used to when I go to perplexity or to chat GPT or whatever? Like, what is the app layer?

Starting point is 01:40:16 Yeah, that's how it looks like pretty much at this current point. And I expect this space to kind of like blow up massively. I mean, the VC investing in the traditional AI world has been huge this year, to say the least. So the concept here, is we talk about in crypto this fat app thesis. I think this is the realized version of what it's going to end up becoming, right? So human-facing or in many cases, agent-facing apps will dominate online interactions. And I think it'll funnel more people into crypto because you won't even know that crypto is being used. I can actually give you an explicit example. So I met this team. I think they're called Taoshi.

Starting point is 01:40:58 they're a subnet on Bitenser, which is this, you know, this AI layer one. I met this team at permissionless, actually, guys. And I spoke to them and I said, you know, what do you guys do? And they explained that they basically said that their service produces financial trading algorithms for small hedge funds. So these are traditional hedge funds that are trading like very traditional commodities, right? Not crypto, no crypto assets inside, no meme coins in sight, right? And I asked them, well, okay, you know, how are they paying for this thing? They say it's in USD or USDC.

Starting point is 01:41:37 And I'm saying, okay, do these guys have any idea that there's crypto rails in the back end that are using these things? And the simple answer was no. I then kind of like dug into the archetype of this team. And none of them had worked in crypto before bar one of their co-founders. And this is a 50-person team that are selling, you know, six to seven figures per year to these traditional hedge funds. So I think it's pretty astounding to talk about a world where we want crypto rails. You know, you talk about the defy mullet back in the day, where people have no idea that crypto rails are being used in the back, being realized so soon and so

Starting point is 01:42:14 early on in an emerging sector within crypto. It's pretty awesome. Huh. Huh. That's pretty cool. Okay, so with all of these stacks, we just ran through them, the foundation middleware app layer. when you just talked about like the fat app thesis is that is that where where are vCs allocating is there is there is there a debate is there contention about where value is captured because this was like my original conversation getting into crypto is like what part of the stack captures the value is honestly why why we started bankless is to kind of answer that question at least in the crypto world like Ryan I came to the conclusion that money is like where a lot of the places is what the where like value aggregates which would allude

Starting point is 01:42:55 to like somewhere down at the bottom. But I have no idea if that extends itself to like this crypto AI stack. So that's the question to you is like where does value get captured? So I'm going to give you the right curve take to begin with. And then I'm going to go, I'm going to lean pretty hard into the left curve take. And so the right curve take is the sophisticated spectacles take. VC coins. The VC coins, yes.

Starting point is 01:43:16 And then the left curve take is the 70 IQ. Don't think about it. Correct. Correct. So traditionally how we've approached investment within the crypto sector in particular is infrastructure. We need rails to be able to even do anything. And I would argue that over the last 10 plus years, we have made amazing progress. We now have stacks that can have huge amounts of block space versus transactions, millions per second, et cetera, et cetera,

Starting point is 01:43:44 you know, L2s, L1s, et cetera. So I actually think that we're at a point where we have a performance capability to do a lot of these things. Having AI agents will actually increase that requirement by 10 to 100 fold. Why? We are going to need more crypto infrastructure with AI agents.

Starting point is 01:44:03 I think there's going to be a larger demand on block space when we have AI agents because they're going to want to make more microtransactions than the average user does currently during a peak bull phase, for example, when we're all trading shit coins, meme coins or whatever coins you want to

Starting point is 01:44:19 describe. So from a fundamental infrastructure layer, we need to be able to create infrastructure that is optimized for these AI agents. Now, I'm not suggesting necessarily that this requires a new type of L1 or L2. It just requires more added thinking as to how this stack is built and how it integrates with our newfound user, which is an AI agent. So with that being said, there's a lot of investment that I'm seeing, at least on the VC side of things, in infrastructure L1s and L1s.

Starting point is 01:44:53 twos, particularly, like how to build all these different things. We're seeing a lot of these agents and crypto projects launch on base, for example. And, you know, there's a reason why they're launching on base versus not on ethel1, right? Now, if we look at other layers of the stack, from a kind of right curve perspective, we need different forms, or sorry, different networks to be able to aggregate data. We need different networks and architectures to be able to figure out what the best way is to serve compute and aggregate compute. These are the infrastructure investments that I think are going to underpin a lot of app development in the future.

Starting point is 01:45:32 So we could just leave it at that, right, and assume that, okay, we're not going to see any crypto AI apps for a while now and we should just kind of focus on the infrastructure layer. You know, run back the last couple of cycles and, you know, it's a pretty simple playbook. Here's my thesis as to why I think that's going to be extremely incorrect. I think that Crypto AI's number one breakthrough case is going to be a consumer app. In particular, I think it's going to be an agent. And taking it even to a spicier level, I don't think it'll be a financial agent. I think it'll be a human-based agent that attracts attention and requires engagement from your average X user or farcaster user.

Starting point is 01:46:14 I think it's someone, I'm sorry, I think it's an app that people communicate with pretty frequently in a chat interface that does a bunch of things for them. And it doesn't necessarily need to be financial. I'm getting visions of just like an AI neopet, like a neopet that's just like super smart and actually like, you know, sophisticated and like fun to engage with. Is that kind of what you're illustrating? Similar.

Starting point is 01:46:39 I mean, we can talk about a few examples. So I think the number one thing that drives a lot of financial value in crypto and has proven over the extent of its life cycle has been narrative. And narrative can be distilled into something as simple as attention. So then you could argue, well, the thing that can drive the best attention and narratives will be the thing that drives a lot of value into crypto, a lot of investment into crypto. Now, we've seen a few examples with the Gospel of Goatsy, which is an example of a social AI agent that.

Starting point is 01:47:19 can talk to users and that can reference human slang and sounds completely like a human being, but it's just an LLM in the back end. I'd capture a huge amount of attention. I mean, like this Twitter account has gone from 5K to 180K currently in three and a half weeks, which is pretty insane. And all of those have, you know, tweets regularly 24-7, huge amounts of engagement. So I would say that it could come in the form of that neopet example, David, I think it'll get integrated into games.

Starting point is 01:47:51 You're seeing the prime guys do this a lot. You're seeing this appear in farcaster agents with the higher community, for example, where you can just kind of like chat to it and it's doing a lot of things. You have the financial element as well with AI16Z, which is kind of a play on A16Z, right? So it's like an AI powered fund manager. So I think there'll be a few different instantiations of it, but I think that's where the breakthrough app is going to come.

Starting point is 01:48:19 There's been a couple of themes in crypto over the last, like, 18 months in the, like, despair market that it's felt. There's just, like, rabblings about, like, man, it's been, like, how many years since we've invented crypto? Where are our apps? Like, we don't have a killer app. Sure, stable coins. But, like, honestly, that's just the dollar. So, like, where's the apps has been a big theme over the last, like, almost two years now? Where all the users has been also another big theme.

Starting point is 01:48:46 We don't have any more users. We're actually like starting to like it feels like it's starting to leak users. Like people are quitting crypto. So we don't have any apps. You don't have any users. And then also like meme coins have been a third theme, which has like kind of created the attention economy. Like the attention economy, people have realized that attention is actually just a hugely valuable thing. This is why meme coins have such product market fit.

Starting point is 01:49:07 It's because they drive attention. And so I'm seeing like a lot of these themes that we've had like define crypto over the last two years. lack of users, lack of apps, and then the attention economy. And AI agents kind of check all of those boxes. Like, well, we get new users. They're just AIs. We get new apps. They're the agents. And how do they work? They drive attention. And so there's like a big void here that I think like the whole agent like thesis idea. Again, this is just an idea, but a lot of people pointing towards it is like filling all at once. Mm-hmm.

Starting point is 01:49:43 EJAS has been great. Thank you for walking us through the AI agent thesis. Thanks for having me. Bankless Nation, you guys know the deal. Crypto is risky. Mean coins are risky, no matter if there's a human or an AI agent chilling them. This is the frontier. It's full of robots and AI overlords, and it's going to look really weird.

Starting point is 01:50:01 But, hey, at least it's exciting. We're glad you're with us on the bankless journey. Thanks a lot.

Bankless - How Crypto AI Agents Will Take Over the World | Ejaaz Ahamadeen

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.