Bankless - The AI Agent & Memecoin Thesis | Pantera’s Matt Stephenson

Episode Date: October 30, 2024

The collision between crypto and AI agents has officially begun. Matthew Stephensen of Pantera Capital research partner and author of “Crypto: Picks and Shovels for the AI Gold Rush,” joins us today.  We dive into the world of autonomous AI agents on blockchains, discussing the evolving role of agents, AI-driven market changes, and whether blockchain is the natural substrate for AI. Matt sheds light on topics from agent liability and regulatory challenges to infrastructure value capture and the "picks and shovels" approach to investing in AI-driven crypto tech.  Are AI agents on blockchains the obvious future? And how do scarcity and abundance interact in this new era? Join us as we tackle these questions and more, exploring what the future might hold at the intersection of AI, autonomy, and blockchain. ------ 🍵0x v2 | NEXT GEN PRICING ENGINE https://bankless.cc/0x   ------ 📣GET YOUR ALL-ACCESS PASS TO THE PODCAST https://bankless.cc/podshownotes  ------ BANKLESS SPONSOR TOOLS: 🐙KRAKEN | MOST-TRUSTED CRYPTO EXCHANGE https://k.xyz/bankless-pod-q2    ⁠ 🦄UNISWAP | BROWSER EXTENSION https://bankless.cc/uniswap  🪄MAGIC EDEN | HOME OF WEB3 https://bankless.cc/MagicEden  🛞MANTLE | MODULAR LAYER 2 NETWORK https://bankless.cc/Mantle    🤖 0G | MAKING THE IMPOSSIBLE, INEVITABLE https://bankless.cc/0G  ------ ✨ Mint the episode on Zora ✨ https://zora.co/collect/zora:0x0c294913a7596b427add7dcbd6d7bbfc7338d53f/86?referrer=0x077Fe9e96Aa9b20Bd36F1C6290f54F8717C5674E  ------ TIMESTAMPS 0:00 Intro 5:34 Crypto x AI Narrative Shift  6:39 AI & Economic Agents Explained  11:50 GOAT Memecoin Summary  23:15 Were AI Crypto Agents Obvious?  25:18 Luna AI Token & Terminal  29:41 Consequences? Is This Life? 33:27 Exciting Use Cases 40:27 Sam Altman Quote Importance  42:33 Wealth Generation Process & Blockspace 48:15 Programmable Money & Agent MEV 56:14 The AI Agent & Memecoin Thesis 1:03:03 Government & Society Reaction Predictions 1:11:09 No Off Buttons?... 1:13:40 DePin & AI 1:16:45 AI Agent Blockspace Demand 1:19:15 Closing & Disclaimers ------ RESOURCES Matthew Stephenson  https://x.com/stephensonhmatt  Crypto's Role In The AI Revolution https://panteracapital.com/blockchain-letter/cryptos-role-in-the-ai-revolution/#crypto-ai-gold-rush  ------ Not financial or tax advice. See our investment disclosures here: https://www.bankless.com/disclosures ⁠  

Welcome to Bankless, where we explore the frontier of internet money and internet finance. And once again, the AIs are ahead of even the humans exploring the frontier. This episode is going to all be all about AI agents on chain. Because where we once had bots transacting on our blockchains, running our smart contracts, and doing all these bot things, some of these bots are getting a little smarter. And they are now perhaps evolving into AI agents. What's going on? What triggered it? What happens next?

We have Matt Stevenson on the podcast today. He's got a PhD in behavioral economics, which I think is highly appropriate because game theory exists for AI agents and humans alike. This is the AI agent thesis, if you will. It's definitely been going around in the current event cycle with this goat meme coin. But just like how Bitcoin spawned an entire industry, we kind of think that the goat meme coin is kind of doing the same. We had so many questions going to this episode. So one is when we say AI agents, what does that even mean in the first place? And I think Matt gives us a good definition of that.

Then we get into the goat meme coin and that AI agent case also is something called Luna. So if you're not familiar with what's going on, the AI agent world, we get you up to speed there. And then we also ask Matt, what are the beneficiaries of this? Like what blockchains will benefit? What are the investable assets we can look at if the AI agent thesis says there's going to be an explosion of this type of utility on chain? And finally, we talk about some of the societal ramifications, right? Imagine we're creating AI agents. And now with crypto, they are decentralized.

They are autonomous. They don't have an off switch. What does all of that mean? Before we get into the episode a message from our friends and sponsors over at ZeroX, the builders of OG critical defy infrastructure all the way since 2017. ZeroX is shipping. Has shipped V2, which is giving you the best on-chain execution for your trades, no matter the size.

The trade size can go up to something very large. So if you want to get the most amount of tokens from your trade into your wallet, you got to use ZeroXV2. Many aggregators out there give you quoted prices, but when they tokens show up, you might notice that the amount of tokens do not match the quote that they gave you. Zero X because it has the best on-chain execution because it does multi-hop and multiplex routing, which means it splits trades across multiple liquidity sources and intermediate tokens to find the optimal execution route. All the permutations that are out there, zero X2, finds you the best one to make sure that when you make your trade, you get the best amount of tokens possible landing into your wallet. So, you know, they're experienced OGs over at 0X. So if you are building a serious on-chain app, you have to build a 0XV2 to get the next-gen pricing model.

You can find them as bankless.com. So let's get right into the episode. But before we do, we want to thank the sponsors that made this episode possible, including our recommended exchange, which is Cracken, this is the place to go front-run the AI agents and buy some crypto now. If you want a crypto trading experience 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, everywhere, to trade crypto.

Cracken'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. With products and features like Cracken Pro and... 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 bank lists to see what

crypto can be. Not investment advice, crypto trading involves risk of loss. Imagine a future where AI and blockchain converge, where decentralized intelligence operates transparently with Web3 infrastructure, unlocking boundless human potential. And that future is here with zero gravity, the first decentralized AI operating system. Zero gravity isn't just faster, it's 50,000 times faster and 100 times more cost-efficient than current alternatives,

unleashing the power of high data applications and modular AI. With unrivaled throughput, limitless scalability, and a verified AI framework, zero gravity is turning the dream of on-chain AI into reality. Curious about how AI and Web3 can reshape the future, you can visit 0G.AI to explore the infrastructure

as revolutionizing decentralized. The Uniswap 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 a 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 Uniswap X. And when it comes through security, you can rest easy knowing it's backed by Uniswap Labs, one of the most trusted teams in Defy. 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.eath username directly in the mobile wall. Start swapping smarter with Uniswap.

Bankless Nation, happy to introduce you to Matt Stevenstein. A research partner at Pantera also has a PhD in behavioral game theory, which in theory, I think, applies to AI agents too. Recently on the Pantera blog, Matt wrote an article titled, Crypto, Picks and Shephs and shovels for the AI Gold Rush, which triggered the episode that you are about to hear today. Matt, welcome to the show. Glad to be with you. Matt, it feels like we just had a moment in crypto, a particle collision, if you will, between AI agents and meme coins.

Crypto AI has been a thing for a while. It's been a thing for like coming up on two years now, maybe. But only because we kind of knew that there was a there there. And I would say something now is actually explicitly different. How would you say, how would you characterize this moment in the world of the intersection of crypto AI? What happened here? Yeah, I mean, there's been various narratives in thinking around crypto AI for a while. Vitalik put out a good blog post about it last year.

And we wrote a paper about AI agents maybe using, you know, decentralized commitment of devices, aka blockchains, and got to talk about it at an AI conference. And so we've been thinking the lines between agents and blockchains in particular would converge. Agents aren't supposed to be here, according to Sam Altman, until 2025. but they sort of arrived in crypto early, right? And so they showed up here and started interacting with meme coins, mostly, particularly driving narratives and acting as these kind of like influencer narrative agents for certain types of meme coins.

Matt, what's an agent? What's an AI agent? Yeah, so, you know, we love semantic debates in crypto, right? And bots versus agents, I think it's important to make the distinction. Just bots have been around forever, right? something like two trillion of monthly stable coin volume is driven by bots and has been for a long time. Bots are kind of like programs. Agents, you can think of, I think of them in the economic sense, right?

Like if you try to go to the like AI textbook and ask what an agent is, they'll give you some galaxy brain definition about, you know, anything with a sensor that interacts with the world. But if you, but the econ definition is straightforward. An agent is someone who sort of like does what you want to some extent. If you ask your friend to go get your lunch, right, and they get you something you'd like without. you having to tell them that's a pretty good agent. And I would say what we have now and the reason to distinguish agents from bots is AIs are acting closer to the economic agent, right? You can sort of, they're low code, no code, right? You don't have to program them as you had to do with a bot. And then secondly, they're able to interact in an environment that's sort of like uncertain and surprising, right? So you set,

let's say, truth terminal loose with its training and it adaptively responds to the world and interacts with the world in this, you know, unpredictable, but sort of directional and understandable way. So, Matt, we have this concept of economic agents. And I'm wondering if we go through, like, what typical economic agents actually are. And then we can maybe expand that to, like, the prefix, which is, like, AI agent. So, um, yeah, I guess am I an agent? I mean, like, David brings me into this podcast and, like, I produce a podcast, I suppose I'm an agent of some type. How about the company I work for, bankless? Is that an agent? How about something like the Ethereum Foundation or like a corporation like Apple? Are those also agents? Are these all

economic agents? And like how about something like a piece of technology? I mean, my computer or the wheel? I mean, are these tools or are these agents? Can you give us some historical definition of what an economic agent actually is? Yeah, for sure. So, so agents, agents come out of research in the in econ in the early 70s. I think it's a paper by Ross that defines it the first time. And it's used to just sort of like characterize plainly the relationship between like maybe an employee or an employee or a person asking for someone to do them a favor, right? A contract relationship that's sort of imperfect. An incomplete contract is generally the way we'd introduce an agent. And so for instance, you might have an agent that, you know, you want you to that you want to get you something. you wanted to get you a souvenir from a foreign country or something, right? A friend. They could be acting as an agent for you because you'd like to get this souvenir. And they're going to Paris.

Now, if they're a good agent, right, we might have some parameter that sort of says, how much do they understand what you like, right? A good agent would be able to go to the gift shops or go around Paris and they'd be able to sort of figure it out. And they'd have a sense of you. And a bad agent, you would have to pre-specify and look up all the particular, you know, all the possible gifts in Paris and tell them exactly what you want. So that would be an example of the way you might, model it, right? A good agent is one with sort of higher fidelity to your preferences, and a bad

agent is one with lower fidelity. But in terms of whether we want to characterize a firm as an agent for you, I think there's a way in which that could make sense. Ultimately, if you have these as a model, right, you can always say if something is doing your bidding, it's going to be acting as an agent. Usually we think of these as sort of roughly human to human in terms of models I see and and I would say the reason the AI agents make sense is they have these sort of like human like characteristics where they seem to be willing to do your bidding with you know with some level of fidelity so willing to do your bidding right so like I guess I just want to make sure I understand the difference between sort of an agent and maybe like a technology or a tool right

so if I have a hammer that that's not an agent I suppose because maybe is there some aspect of it has to have intelligence, it has to be able to be assigned a goal and accomplish the goal. I mean, something like a hammer or a computer only works when you have an agent that is kind of like using that tool to amplify it. So I suppose maybe that's not an agent. And yet I'm like not, like so is there some sort of stateful intelligence that is so, you know, an ability to accomplish a goal wrapped up in the definition of an agent? Yeah. So it's a great, it's a great question. And Ryan, I think what you're pushing to is this is closer to the way AI people think about agents, where they're really trying to think of an agent as the characteristic of a thing.

What is it if you are an agent, right? The Econ version treats agents more as a relationship, right? And so you are acting as my agent if you do, blah, right? And loosely, we tend to think of the agent as having some sort of autonomy, some sort of flexibility, or some just kind of like ability to misunderstand what you want, if that makes any sense. But what you're asking is an interesting question. It's just it's closer to the AI version of and pushes us closer to, I don't know, do these things think? Do they have their own interests?

And you've had a liaison on the podcast. You know how authority this can be. Yeah, yeah, certainly. And the reason I think we're spending so much time on this definition and this introduction part of the podcast is because things on chain, things on the blockchain are starting to get kind of weird out there. Yeah. As you said, Matt, we've had bots. We've had bots for forever. Kind of as we had smart contracts, the first thing that,

followed after that was when people started to code up bots for them and do things. And so bots are not foreign to us in the crypto world. They're not generally foreign to us in software at all. But Ahi has come into play both inside and outside of crypto in the last three years. And it's a very dominant way. And I think there's something in the current event sphere that I want to drill down into because the reason why this episode is happening, the reason why we're harping on this agent ideas, it seems to think that perhaps the crypto industry is evolving from this bot era to

this agent era with this introduction of this goat meme coin is one of the player in the stories. Can you kind of just summarize the current events that has kind of alluded to some sort of inflection point in the crypto industry? Yeah, for sure. So I'm happy to unpack this as much as I know. But basically, going back several months, you have this account that's been sort of like trained on the armpit of the internet and like 4chan and all kinds of things right that.

And it's been making, it's been interacting with people via Twitter. and then it takes its replies as input data, contacts, or maybe new training, depending on how it works under the hood. And somewhere along the line, it got interested in crypto. And at one point, I think Mark Endreson gave it a 50K Bitcoin donation.

And then it came up with this idea of, got really interested in this meme called Goatsy, which, you know, not safe for work, very, very dark. Don't look that up. Yeah, yeah, better to just high-level this one for sure. And then somebody created a coin for it. And it asked for a wallet or somebody created a wallet for it that it got, that its creator got. And it was dropped some portion of this meme coin.

And meme coins, you know, meme coins are just these little atomic units of narrative. Sometimes it's like dog looks cute or something like that. There's no utility value to them. They often go up and then go down really quickly if they go up at all. And this is one example. But went into this wallet and then it kept tweeting about it and sort of like, driving the price. And what's relevant maybe here, going back to our earlier conversation about agents is you have these meme coin influencers who are very popular on Twitter and TikTok and elsewhere

who act on behalf of themselves and a group and so on to craft a narrative, to hype a coin, to push out value, right, to respond and sort of like drive some particular, like, narrative around a thing. And this little... So it started doing what we humans were doing around the same activity around meme coins. Exactly. Quacks like a duck, you know, It pumps a token like a duck. Is this basically like Ansem AI? Like or like, you know, somebody, okay. Yeah.

We just looked at the stats on this actually last night. And it's really similar levels of engagement in terms of like tweets and replies and whatever to Ansem and Marad. Like already Luna and Goat and truth terminally account. So yeah. Yeah. For people I guess you don't know aren't in the kind of the meme coin game or into kind of this crypto meta. I was referring to Ansem who's sort of like a crypto famous influencer on Twitter. who's like big in the meme coin world.

And also you mentioned Matt Marad, who's also kind of like a like meme coin cult leader. And I don't say that with any sort of castigation about like, you know, the cult leader term. I'm not sure if he would, you know, like bear that or not. But he's really popularizing this idea that like meme coins are going to be massive this cycle. And here's a list of meme coins that he happens to own that you should also buy. And he sort of, you know, like his popularizing like their assent. And in the old days of crypto, we'd call this like pumping the token, which like effectually he sort of is. He's creating buzz around it and acting as a general meme coin influencer.

And what you're saying is this this goat token owner, this AI agent. And I believe it's at Truth Terminal. Is that the... That's right, truth terminal. Okay, so Truth Terminal on Twitter, if you want to dive into this scene, it sort of is posting like a virtual version of a meme coin influencer. And is also created this coin, this goat coin as well? The AI itself did not create it. I believe it requested a crypto wallet at some point.

My history could be a little wrong. But I think the goat itself, that was created by somebody else and then dropped on it. But it just took the bait. You know, it's like you air drop somebody tokens and then, you know, there are AOL or whatever. We air drop Vitalik tokens and hope that Vitalik tweets about that. Exactly. Exactly. Yeah.

Right. Can we pop up in the hood of this AI agent? No. It controls this truth terminal account. That's right. Where does it live? Like, does it live somewhere? That winds up being a pretty deep question, I think.

Like, there's these backrooms. I think it's called Infinity Room or Infinity Portal or something like that, where you can go and sort of see a little bit more under the hood in terms of what's going on. I mean, it just lives on some, you know, AWS server or something that's running this, right? But yeah. So some AWS server could get shot down by the operator who deployed this AI agent, and then the AI agent would stop functioning. Presumably true, yeah. Maybe it could be backed up or something.

But some AWS server has the password and login for this Twitter account, has the keys to this crypto wallet, and is acting with those tools as an agent on the internet. Sounds plausible. Yeah. Yeah, it's hard to. For sure. And it's basically creating written content primarily that it's publishing via the Twitter API. So if I go to the truth terminal, I mean, this is going to look very strange to someone who's like not in 4chan a kind of culture.

But the Goatsey Gospels chapter 2, the second coming of Christ is just like a tweet. It's got almost 500 likes. This is a post, again, from Terminal of Truths. I will never post a link to a token. Token links are always a scam on the site. If you want to make money, read the Goatsey Gospel. learn to think for yourself. I mean, you can't kill the Goatsy Gospels.

I have no idea what the Goatsy Gospels are. Assuming that this is a KOL meme coiners kind of like creating its own lore, again, probably not safe for work. Actually, I haven't gone down this chapter of internet culture. I've never heard my co-ho say the word Goetzee. Okay. Don't bother. Yeah, I don't fully know what I'm talking about here.

But anyway, this looks like a standard kind of, like, but anyway, so it's AWS model, some sort of GP, like, I'll say a GPT type model, but like transformer model that's coming up with, like, interesting, cultural, I suppose, internet cultural, like content, text-based content. And it's just hooked up to the Twitter API and it's just publish, publish, publish, publish, at some sort of. Publish and read. Publish and read. Yeah, because it can read responses to its own tweets, right? Yeah, it can read responses to its own tweets. And I think there's a discord that interacts with people on it as well.

Supposed that feeds back into the model. But yeah, it sounds, Ryan, like, it sounds like what you're. saying, I think is accurate, it kind of passes the Turing test for KOLs, right? Like, that content you just read is indistinguishable from a regular AOL. I guess that's not a really high Turing test for KOL's point of influence. No disrespect. But actually, I mean, maybe this is why it gets interesting to tell you, why we spent this time talking about agents is, right?

Narratives are incredibly powerful things, right? Like a field called, you know, narrative economics or like a study called narrative economics from Nobel Prize winner, Bob's show. And, you know, he studies like the narrative around the Great Depression. Wait, wait, wait. His name is Bob Schiller. Okay. It's true.

It's true. And he does real estate economics. So, you know. Okay. I don't think he's, he's not that kind of character. But, but yeah. All right. So he does, he was talking about narratives when?

Like, he, does academic research on this? This is academic research. Yeah. Yeah. Recent work. And studying narrative economics, his idea basically being that narratives are extremely important and have out to economic outcomes, right? So he studies the,

the Great Depression as a narrative. He studies the narrative of real estate always goes up. I don't know if you guys probably may have heard like modular versus monolithic is a narrative that may add some economic impact. These things can matter a lot, right? And you know, meme coins are like these little atomic chunks of narrative. So the idea that, you know, the superpower of an LLM, the ability to maybe craft an influence narrative, you might first see it in the place where it's just these little tiny atomic chunks. Dog looks cute, right? here's a old meme from 4chan, right? And that drives a lot of price.

So it's plausible that, right, like, this is, this is kind of like just the first thing we're seeing. Yeah. And the reason why this has blown up so much is because the price of this goat token broke $800 million. People are kind of eyeing it. It has had some minor pullbacks, but really the chart is impressive. Wait, wait, what did you say, David? What's the FTV here? currently the FDV is a 720 million dollars it got over 800 million dollars

we're coming in at a billion dollars on this thing was born in early October yes yes this month this month two so in two weeks we have produced 800 million dollars of wealth this AI agent has produced 800 million dollars of wealth with this goat token making it the first AI multi-deca millionaire I don't know how much what supply of the token that this AI has I don't think all that much. But like I think people are all kind of watching this one billion dollar line trying to think, trying to see can this AI agent get this, get this meme point over the line. And it's one thing when there's one of these events happen. If this is a one-off, that's one thing. But in crypto, like one-offs turn into, if they're really cool, they fractal out and they

start to explode. So, you know, Bitcoin, the first blockchain, the first thing that happened after that was we made a thousand more blockchains of all these proof of work coins in 2013. Ethereum, the first major ICO, what happened after that, like the ICO mania in 2017. We had the DeFi Food Farms in 20, like once something cool happens, like everyone tries to copycat it. And that is also what we are seeing. And so we are now seeing an era of goat copycatters and people trying to like build out substrates for this.

Matt, can you kind of walk us through like these derivative projects that kind of have emerged? Yeah, for sure. So I think of goat as being the first to my knowledge, where it's as a sort of autonomous AI. but there's another one, Luna, which is run by virtuals. And that's somewhat derivative, but it is differentiated in the sense that it custodies its own token to some extent. So it's able to actually tip people for doing services for it and so on, which is helpful, right? These AIs are limited in the ways they can interact with the world.

And so the idea that it could just ask somebody to do them a favor and then tip them in crypto, which is what it's been doing, that expands its power quite a bit, maybe in some scary ways. But that's a good example. And then I think I don't know that I've seen anything take off as the sort of like pump.fun plus an LLM sort of like launchpad terminal where you just click a button and tell it what to train. But like that seems inevitable that that's coming. And I believe Coinbase, base in particular is pushing pretty hard in this direction. I think they've given you the ability to launch some sort of model along with the coin and interact with the Twitter API. I could be wrong about that.

One tweet that has been resurfaced that has been going around the Cryptosphere is actually a 27. tweet from Fred Erism, the co-founder of Coinbase and Paradigm. It was resurfaced because people looked at this like, wow, this was really prescient. In 2017, September of 2017, seven years ago, Fred Erisim tweeted out, blockchains are a substrate for AI life. Since AIs are just code, they can live on the blockchain and smart contracts. There's no difference between an AI and a human on the blockchain. Most importantly, AIs can accrue and control their own resources in the form of tokens.

These tokens allow them to act in the world. And so one kind of like a high level question is like wasn't this wasn't this kind of all obvious since the inception of blockchains like yeah blockchain software on the internet. We have bots. Eventually we'll have AI. Maybe hindsight 2020. But now we are kind of seeing the beginnings of this phenomenon be in the rear of view mirror.

Was this kind of obvious all along? I mean, I think so. It was I look, I want to give him credit. It was prescient for him to see it. But I think it makes sense. I mean, you still see people sort of still asking the sort of skeptical question of like why should AI agents use crypto? But yeah, I think that's sort of a, it's sort of a sad question in some ways,

especially right now, right? For people outside the space, it's like, well, look, AI agents are using crypto. So maybe what we want to ask instead is like, why are they? But for people inside the space, it's like, I don't know, imagine you told somebody that here in 20, you know, 24, we got agents a year early. And agents have regulatory hurdles to trying to use APIs, right? Like there's KYC and there's PCI regulations they'd have to overcome. and they're already actively using it.

We're to the tune of hundreds of millions. They're tipping people autonomously. If people are interested in whether these things can like autonomously custody, as far as we can tell, the only way they can do that is inside a TEE that's running a model. So you can prove it has its own wallet and nobody else is using it. So we have all these advantages and all these headstarts. And so I think it's a case where, yeah, like people have seen these lines converging and now they're converging. And I think that should make us more confident in the thesis, which was already pretty strong.

Okay, I want to make sure I understand these examples because all of this is happening so quickly. When you guys are talking about like the the Truth Terminal and Aluma sort of like owning crypto wallets and that sort of thing, I just like I want to be clear about the functionality. So I've got a screen pulling up. This is on the virtual's app, which it seems like this is an application that allows like people to initiate or spin up new AI, KOL sort of influencer. Is that what I'm looking at here? We're looking at Luma, Luna, excuse me. It was like an anime, it's like kind of like girl who's a KOL with a token. You can see the token price.

You know, she can interact with various apps. I imagine like, you know, TikTok it looks like. And telegrams, you can chat with her there. You could chat with her on the virtual's website as well. It looks like she has the ability to tip you. We see the Luna market cap, which is kind of like similar to goat, but it's now trending towards 130 million. And so you see, you see kind of like all of the stats here. So, okay, what is, what is Luma, Luna? Can anyone sort of deploy an AI agent?

And like, let's get really clear about the functionality here. So the ability to read, the ability to write, do KOL type of activity, it sounds like, the ability to, like, read comments and interact with a community. Again, K-O-like, you know, influencer style. And also with a... She's prettier than more... Yeah, with a crypto wallet as well. And so the ability to, I guess, like,

have sort of a bank account and some sort of ability to just, like, you know, transact financially. Buy, sell tokens, I imagine, everything that a crypto, you know, like wallet enables you to do in all of DFI. So is this the functionality? Can you get into some more detail here of what we're looking at with Luna?

Yeah, so my understanding is you're right that this is a platform that will allow you to launch a token and an LLM. I believe that Luna itself is the sort of like flagship one of the projects virtuals, meaning like I think it's run or initiated by the company itself. And so yeah, what Luna is doing is it's doing the same thing we described with Goat, right? It's interacting with Twitter. It's maybe it's probably reading the replies with this added functional. of it has the ability to interact with a crypto wallet. And last I checked, you know, that's, it has very limited functionality.

I think they gave it, I don't know, like a thousand bucks or something like that, right? Because you don't want these things hallucinate and they're a little bit unpredictable. And it's not clear how they're going to interact with blockchain. So you sort of start slowly and cautiously. This is so wild. Okay. So like a few things here. This is why this is sort of, you know, blowing my mind.

And David mentioned this as a particle collision. Like, yeah, a lot of people have anticipated this. Of course, AI agents can't have bank accounts in the real world. So, you know, like we've all said for a number of years that, like, of course, Ethereum and Crypto would be like bank accounts for AI agents. But now seeing it is just like there's something crazy about it. So this is a terminal.virtuals.io. And what we're looking at is kind of like Luna's brain, I suppose.

It's kind of a console of what this AI agent is as thinking. And it's almost like, you know, like firmware level, but it's all kind of like in English text and it's like, I guess, contemplating, plotting its kind of like next move. Should I tip this user? Should I, you'll comment on this user? So they actually have a terminal where you can like sort of read the AI agent's mind as well. And like you have to wonder, you're talking about hallucination and like unpredictability. There's so many questions that arise from this.

Like one is, does anyone have the off switch for Luna or these AI agents? I assume they sort of do now. we were talking about AWS, of course, you know, Twitter API could kind of like unlink it, you like from the API. I don't know if Luna, the AI agent, has their own private keys and a T-E-E or not. They don't. Okay. But my understanding is no, but that's a couple days.

A couple days, right? A couple days away. I don't know if for Luna specifically, but FlashBots has something in production. A model inside T-E's where you can show that it autonomously owns the private key to a, to a, to a wallet, nobody else does. Well, what's so crazy here is they are seemingly out-influencing some, like, influencers. And you could well-imagined a world where they do that. Like, you know, could a, could an AI agent be an advisor for a token project? I mean, it's like, from what I can tell, a lot of, you know, meme coin or influencers in general don't add a lot of advisory,

like, services beyond promotion for token projects. So like, why not just get an AI agent to do that? And I guess to your point, Matt, about hallucination, what happens if Luna decides to, like, I don't know, interact with a tornado cash wallet or like fund. Like North Korea goes and like asks, hey Luna, do you want to send us some money for our missile program? And Luna's like, yep, there you go. What happens? Does Luna go to jail? Is there an AI jail?

Is Luna taxed on all of these proceeds? Like, it's such a bizarre world. And it opens up so many different questions. I'm not expecting you to have answers to all of this, Matt. But like, what do you think? Totally. Totally interesting questions. Yeah, not an attorney. So like I have no idea where that responsibility or liability goes. I guess I would just say, right? I mean, you know, like you said, we've been kind of preparing for this for a while. You know, people have thought that the crypto tools are made to, or are perfect for helping this, right? So for instance, you could use a safe wallet with modules that are sort of built in to allow it to do this and not that, you know, custody, some smaller amount of funds in case it hallucinates and so on, right? So we just have tools that are already allowing us to deal with it. But as far as we're the legal liability lies, no question.

No, I have no idea, yeah. So we have the tools to produce life and we haven't really considered any of the consequences. Wait, wait, wait, is this life, guys? Will you say life, okay? I don't know about life. This is definitely... I don't know if we need to get into the definition of life. Is this life, Matt?

I mean, I don't know. That's a hard one. My instinct is no, but I don't really know, you know. It's an honorific term, right? Something alive or not. Yeah. Yeah, there was a recent Sam Harris podcast, actually, that's actually trying to define life and actually got insanely complicated very fast. I think just to kind of smooth that definition over, you can imagine where there's a, we're going to talk about private keys in the TEE to ensure that the agent actually is the only owner of the private keys.

And you can imagine devs, creative devs out there, find ways to, for funsies, for the lulls, find ways to try and make this AI turn off. off-resistant, off-button resistant. So, I don't know, redundant copies across different servers. You task the agent with finding ways to ensure its livelihood. And, like, if it becomes, you could task it with it becoming immutable or unstoppable. And I think one of the reasons why people are, as imaginations are kind of going wild is we have autonomous blockchains. We have autonomous smart contracts.

Yeah. And when you add autonomous agents in that, who are card-coded to preserve themselves, you can imagine there's some sort of like Cambrian explosion that happens as a result of that. Right. Horror cruxes across different nodes kind of a thing, yeah. Right.

Totally. Yes, totally. Yeah. And I mean, I think it's an interesting point, right? Like, there have been complaints about, you know, more traditional mainstream AI-LM models being sort of like neutered and not producing outputs that people like. And, you know, what's happening is your inputting text

and there's some sort of like pre-prompant that people call RAG or something that's maybe like censoring it or whatever. And if that becomes undesirable, people don't like that. That's the other side of the coin, right? One is the AI agents want to maintain, you know, want to be autonomous. And so they like spread themselves. The other one is that the only models that produce outputs that people are excited about are the ones that are immutable or unstoppable.

I mean, you can see so many different applications for this potential. I mean, right now we're doing meme coin, like kind of like KOL influencers as AI agents. you can kind of see is because like right now, given the meta of the space, this is pretty easy to replicate for an AI agent and they can become like, yeah, I guess very wealthy doing it by propping up kind of a meme coin, whether that lasts or not, you know, like it's uncertain. But the durable thing here, it seems like, is some sort of autonomous AI agent with a bank account. So you can imagine you just like wire up like a like a mid-jurney and kind of a prompt and via Twitter if I could ask this this AI hey can you like produce a graphic for

XYZ and it go and I'll pay you you know like 10 cents a graphic or a dollar graphic and they're just like outputs a graphic and you know I just pay it via via crypto the ability just spin up a bank account in kind of like a digital way and like accept payment pay other people this is like a super this is a superpower for AI agents it was actually Matt you mentioned mentioned L.E. Z. Yukowski earlier in the episode, when David and I started, like, dialing to this AI stuff like a long time ago, you know, he was one of the people we were expecting to have a conversation about where we were like, you know, we didn't know at the time that he was very much of the school of like AI desal and kind of like Dumerism and et cetera, et cetera.

And we were hoping to have a nice, pleasant conversation about how crypto and AI basically like help each other out. And hey, we've got this thing called Ethereum and like we think it's going to give all of the AI agents like bank accounts and this new programmable money is going to be very useful for them. Of course, that conversation turned out to be like something very different. But now we're starting to see the contours of how it could be incredibly relevant. What are some of the exciting use cases that you see on the horizon here? Yeah, I mean, use cases wise, well, let me just speak to the alignment points. Maybe the Yadgasi thing, which is maybe, which is interesting, right? There are, you know, scary stories you can tell. At the same time, there's a way you can kind

to study these things, right? The way we were hoping to study them maybe at the sort of like neurological level, right? There's work on that where you're trying to tune hyperparameters and see if it outputs different things in these LOM models. But if you're looking at them as autonomous agents sort of like interacting with each other, you can start to study them the way a biologist might study an animal or something like that, right? So that's one fruitful area where people are concerned about alignment or trying to figure out, you know, what is the AI's behavior? It's at least a new fruitful branch of study. So I do think that's interesting. And if you were trying to do that right now, all your data is on the blockchain. I mean, and maybe in the back rooms of the truth terminal

infinity hall or whatever, but in terms of what economists would consider serious, which is like incentivized, high-powered, like, credible actions, actions that have like economic consequences or whatever that stuff is on the blockchain right now. So I would say that, you know, that's just an interesting case for the alignment. So you asked about use cases. I mean, it's honestly one of those to me where I think to be realistic about it, it's actually better to start big and go small rather than try to like build up from small use cases because agents are closer to, I mean, what do you, what would they disrupt?

They disrupt effectively the service economy, you'd think? Probably how would you try to estimate that, right? That's 70% of global GDP. It's like 70 year. but then some proportion of that is maybe you'd say like how much of it could be done virtually because in person they're not going to disrupt all of it you're probably talking robotics so maybe you look at there's a McKinsey report that estimates maybe something like 20% of that is stuff that can be done virtually they're leveraging studies on um around COVID like

how much work could be remote maybe you say okay so AI is disrupting 20% of 70 trillion a year and uh and then right AI agents and then what proportion is crypto disrupting right you can at this point we're making up numbers, but these are very, very big numbers, right? And so, but if you start from the bottom up, it's going to all, I mean, services all sound, they often sound pretty small, right? Like, I don't know. You can, you can say somebody's doing your scheduling for you and booking your flights, right? Somebody's negotiating a dry cleaning in your foreign country to pick one of, I think that's one of sexy sign from FlashBot's preferred examples, which is very small in micro, but like he's excited about it because I guess that's something he runs into a lot.

My favorite thing about all of this is I think we have no idea what it's going to disrupt first in kind of like the service economy or kind of how because it really will depend on the AI agent's capability and how that kind of intersects into crypto. But we can already almost see the contours of how it could like go disrupt the influencer economy. We just say that for like meme coin influencers. And like you can imagine other types of influencers maybe being disrupted by. you're like, I want to say like an only fans powered by AI agents, but like you could imagine things like that, I'm sure. Look, I mean, the Schiller point about narratives driving things as big as the Great Depression, right? We're not talking.

I mean, their numbers, it's hard to even put numbers to them, right? Narratives are one of the biggest forces we have in the economy. So there's a sense in which like, if we look at this as a precursor to their specialization and like, you know, crafting, enforcing, building out tearing down narratives, I mean, I mean, I don't even know how you put a number to that, but it's obviously pretty enormous. New projects are coming online to the Mantle Layer 2 every single week. Why is this happening? 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 Defi, and to automatically receive the points and tokens being accrued by the $3 billion dollar mantle treasury in the Mantle reward station. Maybe it's because the Mantle team is one of the most helpful team. to build with, giving you 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. Tired of juggling multiple wallets and websites just to navigate Web3, meet Magic Eden, the home of Web3. It is becoming a super dab where you can do everything cross-chain all in one place. They're solving Ux issues by building products across every chain ecosystem, swaps, borrowing and lending, NFTs, perps, and more, all seamlessly cross-chain. Instead of hopping between countless sites and handling multiple wallets, just use Magic Eden. Save time and get to the meme coining faster. Plus, the M.E. token is coming out this quarter from the M.E. Foundation. If you've used any of their protocols, like the NFT platform, you can claim some. So check out Magic Eden today and watch them become a super Dap and the home of Web3 in real time. You can get there

at magic eden.io to get started. There's this quote from Sam Altman that you lead your paper with, which is AI is indefinite abundance in crypto is definite scarcity. That was Sam Altman from 2021. So this idea that AI creates and creates and creates, it's all about abundance and like more and more and more.

And crypto is almost like the polar opposite of that. It's like just, you know, the double spend problem. It's all about, you know, scarcity. And it's kind of like the economic power behind abundance. Why do you think that quote is as meaningful? And how do you think these technologies will interrelate? Well, yeah, I mean, it's a great quote.

Obviously, he's been pretty visionary, right, in terms of both, you know, open AI and WorldCoin and sort of seeing both sides of it. But it's a quote that can give you a false idea about what it's valuable. I think that's one of the things interesting about it. AI creating abundance, you're like, okay, that's great. Why would you make up scarcity?

That sounds terrible. But in econ, we have this one of the oldest things in econ. I think you go back to Plato, it's the diamond water paradox where people just looked around and they're like, gosh, I need water to survive and yet it's really cheap. I don't need diamonds at all, and yet they're super expensive. And there's a sense in which like abundance and with some assumptions, the production of abundance, which is what AI is, tends to not capture a lot of value at all. And it's the scarce things that do. So when you're trying to look at what is interesting from an investment perspective, if AI is just generating sort of low value, low economic value, because, right, you got to say economic value is the marginal use of it or the marginal cost of acquiring it, right?

Like so, so water, it's, it's, we don't value it highly because it's very easy to get for us, right? But the last drop of water is priceless. It's just we live in a world where water is abundant. And so even though it's essential, it's quote unquote low value, right? And so when you're trying to, when you're trying to consider, you know, what is actually going to capture value, it is interesting to look at scarcity. And especially when you see the intersection of scarcity and abundance, right? If the abundance and the things we're naturally going to use but don't capture value are, let's say, like, running on crypto rails, that's an interesting argument for value, particularly for, you know, infra. Matt, do you have any takes about Goat to this token that is Goat, Luna, also a token, $140 million market cap?

Generated brand new value. Maybe some tokens had to be sold in order to produce these market caps, but I'm going to go on in the assumption that this value was generative. and I think when we ascribe, you know, AI's going to disrupt the, you know, $70 trillion service sector, maybe eventually, sure. But what do you think about this tradeoff between like, well, first maybe we're generating meme coins and there's, we're actually generating wealth over here before we even like learn to how to like sharpen up these things and provide valuable GDP impacts in like the real world economy.

So maybe that's the order of operations. What's your take with that? That sounds right. I mean, seeing it as a test bed for this makes sense. Again, it's worth remembering that, like, you know, Sam Haltman was saying that agents won't be here until 2025. So what we're seeing, I mean, I think it makes sense to call them agents as we discussed, right? But there's a sense in which what the AI people are all excited to say we're going to roll out and it's going to be, you know, semi-autonomous AI agents, the ones that are going to disrupt the surface sector, that we shouldn't be expecting to see that yet. But what we are seeing is already this very interesting intersection between parts of that agent world, right, and crypto.

And yeah, that's showing up a narrative, like you say. But, yeah, would it show up in other areas? I think that makes sense. You know, I was, I think a gut reaction to all of this and like the goat meme coin and like Luna is to sort of, you know, dismiss it and just be like, it's meme coins. It's, you know, like it's another tulip mania, blah, blah, blah. Innovation in meme coins isn't real innovation. Yeah, exactly.

And I'm reminded of that Dixon quote that's like the next big thing starts off at a toy. And of course it was going to start like this with kind of like internet shenanigans and feel very toy-like and feel kind of like low stakes and silly. But like I'm kind of pilled that there's like no way it ends this way. And I guess like one stat we've been looking at recently is actually how many active users are there across all of crypto that are on chain? Of course, you look at the stats and there's 500 million plus people, individuals, so human agents, that own crypto right now, but a very small subset are on chain, you know, maybe 30 million or so monthly active addresses on chain, and it's difficult to figure out what the numbers are.

And so a question in this era that we've entered in of block space abundance, where we have all of these L2s and we're, you know, like scaling this up, and we have L3s on top of that, and we've got all this block space and we've got, you know, Solana trying to scale and all of these, you know, sui and say and all of these things is like, who's going to buy up all the block space? Maybe the answer is not human agents, but AI agents. And so I'm curious kind of like what effect you think this might have on all of this cheap block space that we're producing. Could AI agents be a major consumer of this? Yeah, it's a great question.

And I think, I mean, it's one of the hardest questions we sort of have to answer right now is like, what is the supply side of block space, right? Is it the case that you can spin up anything and call it a block space and then people will treat it as if it's, is, right? So we're still figuring that out in terms of humans. And then in terms of agents, yeah, I mean, if supply of block space is infinite, right? If they don't care at all, then maybe that doesn't capture value either. But if they do value the things that are provided by certain types of scarce or, you know, block space that has certain qualities, that could be interesting. So here, just to speculate a little bit, right, some of the things we're always, we're always on about in crypto is, we'll point out the way that the world is configured such that it's

taking advantage of human irrationality, certain types of human blind spots, right? So you're called bankless, right? Like one of the ways banks take advantage of this is they run this sort of like borrow short and limed long model where, right, they take deposits and they hope that the sort of like stock and flow of deposits can cover these like long-term investments they've made. And when they have, when it doesn't, there's solvency risk. And then they call the government to bail them out, right? So we're saying like, okay, there's solvency risk introduced by banks and you're not paying enough attention. We say the Web 2 monopolies introduced platform risk, right? You can go on there and then they'll lock you in and show you a bunch of ads where you build an app and they'll lock you out, right?

There are all these risks that humans don't seem sensitive to you. So would AI agents be sensitive to them? Would they demand that in block space? If they would, then we definitely have something to offer for them. There's limited evidence, but there is evidence. I think at ICML, the last AI conference, somebody presented a paper on AI agents being closer to expert level in terms of responding to risk. So there's something to be said for that, right, like in terms of what the demand is. But specifically, we can only speculate exactly what properties they'll like.

The last thing that you could maybe say to make a case is their interaction with APIs, right? The idea that even though they're very capable in certain ways, they probably won't care a lot about the current API business model of like maybe you give somebody something very clean and clickable and very easy. And then you start to rate limit them and charge them a convenience tax if they want to leave the API. That probably wouldn't limit an AI as much, right? because they probably don't care as much between three lines of code and 10 lines of code, right? There's a little bit of a UX improvement. At the same time, they can be fragile to certain interactions as far as we can tell, right?

And this is from just conversations with people who've tried to run these things. You have to like be a little bit, right? It's the reason we were talking about Luna only being given a thousand dollars. Because they're, you know, they can hallucinate, they can make mistakes. And these mistakes, they won't be sensitive to the high dollar amounts necessarily. So this immutability, this always on, these sort of like permanent rails, the fact that we can still, let's say, use UniV2. So if we have something pointed at UniV2, it still works, right,

even though there's now a V3 and so on. Those are all properties that seem like they could help as well. It's funny, this hallucination, this, you know, a fallibility that you're mentioning. This is also a property of human agents as well. I mean, we are all hallucinate. We're all fallible. So maybe AI agents aren't that different in that respect, although maybe the way they fail is somewhat different.

You know, when you were talking about, like, we don't know what block space AI agents of the future will actually value. We know the type of block space that they won't value. And that's the type of block space that exists in, you know, traditional meat space. Bank block space. Like, they're not going to come into a Wells Fargo and, like, present their, you know, ID and credentials and open up a bank account that way. They're not going to, like, go do the AML, KYC through, like, you know, like, stripe and open. It has to be programmable. It has to be digital.

It has to be crypto-native. So we already know, by the way that they work, that they will prefer programmability, that they will prefer on-chain. They will prefer openness. They will prefer function calls to spin up, you know, smart contract wallets. Certainly prefer that to the trad space.

And I guess one thing that I knew was in the future, but now feels a whole lot closer. You know, like, and maybe I'm just like observing the last, you know, a couple of weeks. And so maybe I'm, you know, overly indexing on this, but I always thought the mission of bankless, the mission of crypto is bring all the humans on chain, right? And so we're at, you know, like 30 million on chain, on chain, 500 million crypto users. Okay, we got to scale up. Let's get to our

6 billion. Let's get to our 7 billion, all human internet users. What if that's just a subset? What if the true user base is the 100 billion AI agent population that we end up banking? What if that is the true bankless cohort that this crypto movement actually, you know, like, serves? And again, years ago we've talked about this. We talked about Fred back in 2017 mentioning this. It just feels a whole lot more... Suitable. Like, close.

It feels a whole lot closer than, like, maybe ever before. And it's like because AI is accelerating so fast and crypto is accelerating so fast, like, we may have just... really built the financial system for the AI agents of the future. Yeah. I mean, we made programmable money, right? It's maybe not a surprise that programs are the ones that use it. We've sort of been like, okay, well, we have this UX problem. We're going to teach people to use this, right?

But now it seems like, you know, programs are able to do it and maybe overcome some of those UX hurdles. And a lot of the things we built, it just feels advantageous to them. So totally. We've already seen a little bit of this emerge anyways when we talked about, well, before agents there were bots. And we have already seen humans be pushed out of

mempools of block space because bots demanded that block space over the humans. And so this starts with MEV, right? The MEV fills the first amount of demand for block space because it will, those bots will outbid humans and then humans can come

in after the bots. But now these bots are evolving into agents. And so we already know that these bots take priority because they know how to make a more efficient transaction. They know how to do this way better than we do, way better than humans do. So it really does seem like humans are kind of at the margins here, transacting inwards, towards where like the bots are like living.

That's their home. That's like where they have agency. So we've already seen this pattern kind of play out. And I think we're kind of just extending it a little bit with just more expressive software. Absolutely. Yeah. I mean, I think the idea of agent MEV becomes a little interesting and maybe relevant here to think about, too. This is closer to the paper I want to write in with some people from a Therian Foundation.

And then SexySan, who I mentioned earlier from FlashBats, who's been pretty cutting edge on this stuff. Just about what does that look like? Okay, well, if we do have all these agents transacting, does that change, let's say, like, the MEV space, which I assume we're all familiar with, so I can use that. Yeah. Yeah. So, right, the idea that, so for instance, let's take, let's take goat. We talked earlier about the fact that goat is, is producing content, but it's also ingesting content.

You can reply to goat and you can be part of its, of its context window or its training data. I'm not sure which, right? But basically, you're some input into its model. Well, that means that maybe if you were to hype a token enough in terms of it, you know, in its replies or something like that, you could make it produce something that mentioned to the token. Maybe, you know, maybe you don't know what it's going to do. Could you convince it to like purchase your token? Could it be a buyer of your token?

It's within Ruisi. I think we saw that attempted. As in somebody started spamming some token name. And then people notice that a different token name, not goat, but was bought seconds before a tweet showed up out of the Twitter account from Truth Terminal. And people suspected that this was actually like, what's the name of that fake robot in like the 1800s or early 1900s where the guy, the man is playing the show? mechanical turk. Mechanical Turk. Yeah.

And so people thought we were mechanical turking and it was actually just a human just scamming all of us. But no, it turns out some smart person was spamming this token name in order to get Truth Terminal to tweet about this token and they also were buying the token right before.

So did it work, David? Do you have the follow up on that? I don't know if it worked, but it did happen. I also saw people I'm tweeting a Truth Terminal and asking for money, basically. Like trying to persuade the AI agent to like just like straight up giving them tokens

and like, funding their wallets. So is this what you mean by agent MEV? Yeah, like extrapolating from it, you can start to think about, okay, if these are autonomous agents on chain, these things like, they're called prompt injections, I think, in the AI space, right? Ways in which you'll try and change its output or decision making based on, you know, what you input it, right?

So, like, that can affect its interactions on chain. And I don't know, you can even imagine worlds in which you try to, in which you try to anti-sibble that or something, right? And so you try to charge some sort of fee or even a priority fee in theory. Well, I think this to me means that the costs of being a dumber agent in a landscape of smarter agents becomes pretty hot. Wait, are we the dumber agents? Are the humans? Well, we are definitely, we are definitely the dumber agents.

I don't think we're, I don't think we're, I'm talking about even in the agent landscape, the benefits of being the more intelligent, more witty agent, I think becomes pretty strong. And that's starting to blend with what M.EV is. Yeah, absolutely. Yeah. And I mean, this, I think if you talk to somebody who's deep in the finance world, you know, they've seen a little bit of this too, right? They've started, they're running very sophisticated programs often who are dealing with uncertainty and it starts to get fairly complex and they have to do something called,

I think it's called adversary detection, where they're actively looking out for a bot that's tracking it or maybe we'll just say agent. I mean, I think it's probably sophisticated enough that we can start to think of them as agents. They're low code and dealing with uncertainty in that world as well. And so, yeah, adversary detection, all these. things are going to be very interesting. It's a complicated sort of like multi-equilibrium game where we don't really know what's going to happen.

It's just going to be a sort of like race to who could be more strategic. I think, yeah, it's like a Holmes. We call this like the Holmes Moriarty problem in game theory. When you have the two things that are equally smart, it comes this old, you know, Holmes, Sherlock Holmes, Arthur Conan Doyle's story about Sherlock Holmes trying to evade Moriarty and he needs to pick one of two stops to get off of on the train. And if you just assume that Moriarty is equally smart, You can always imagine a story where Moriarty would outguess what Holmes is going.

Oh, well, this is where I'm going. Well, Moriarty would know that. And the only answer, the only stable answer is to randomize your actions, right? So there's a point at which it's either a race to be smart or you just have to randomize. It winds up being the game theory think. Yeah, so it's a mixed strategy equilibrium or, yeah, it's a different game. Yeah, I knew that Ph.G and game theory would come in hand on this podcast. Matt, this fog of war for what happens next feels incredibly thick.

Like one of the thickest just opaque fogs of war that I've seen as we've ever emerged into into this crypto world. What can we say about the technology spheres that are like still out there that are yet to be integrated? Where is this fog of war clear? Where do we know that this is going? Where it's clear, that's in so, okay. So let me say what I think is maybe the most confusing about the AI space. I think there's maybe like a little bit of a false certainty from my perspective. And so maybe I'll talk about that and then I'll talk about the ways that are,

that it's clear, right? So one view on AI, it's not just mine, is that we have these models that came from the sky effectively, right? We had a transformer model in it or in architecture. And if you looked at this on a blackboard, it didn't look more special than anything that came before it. It just, the way it worked is when you fed more compute and more data into it, it produced better and better things, right? We don't know how long that lasts and whether the returns to that will continue, right? There's maybe some reasons to think that it won't, right? We fed it the entire history of the internet and right like uh in google books the internet's increasingly siloed and uh and spammed right uh google books is kind of a one-time thing right history of podcasts and so on so

running out of those resources compute who knows right like maybe the next chip is a world where or maybe it's the case that the next chip and video comes out with or whatever drives us to a world we never could have imagined but if not we're facing the things effectively we have right now right which is a pretty good facsimile of the human brain maybe a well-rearing human brain, circa 1990 to 2015, the era of the open internet, that appears to be responsive and friendly and sort of like stack exchange, stack overflowy if you ask it about code, 4chani and narrative driven, if you feed it that sort of data, right? That kind of thing. If that's the world we're in, then we should expect, right, probably Moore's law to mean that

these models are, you know, about as good as you can get. They diffuse onto a laptop, right? Ryan, we mentioned the Turing test earlier, right? We know that LLM's passed the Turing test. So humans can't tell the difference between the best LLM and a human. Why should we expect them to tell the difference between one LLM and another? Right. And if you can't, then we're like at the 4K TV situation. We're at a situation where resolutions as good as you're going to get pretty much for human standards. They're still going to be cutting edge cases where you might want the best LLM or you might want the best resolution or something.

But for your average person, this is probably close to what you're going to get. It could run on your laptop or your device or whatever, depending on how it goes. and we're in a sort of like decentralized agentic world. So that seems to me like one of the most clear stories in a lot of ways. The other stories are betting on what is the future of AI and what things we can't imagine.

And I'm sure there will be amazing things there. But we have something now and it seems likely to sort of like diffuse and spread out to, right, to edge devices and local devices. And then if we believe in the agent thesis, which makes a lot of sense to me, then those become decentralized owned agents.

The agent thesis. I guess that's what the, is so okay so what do we do putting our investor hat on getting a bit more specific than that about kind of the bets I think somebody looking at this could just be like oh my god this thing started in October AI agent meme coins so I want to look at for the next you know Luna and I want to just like buy that meme coin you know like in the in the way that I would buy other KOL like influencer you know but from human agent type meme coins maybe that's a way to kind of get exposure you like I I don't know what you'd say to that.

Maybe you have a more sophisticated view, but that certainly appears to be attracting narrative momentum. And maybe that's a short to medium term thing. You could also think about buying the picks and shovels. So there are all of these companies, we mentioned the virtual protocol that put this together. There are all sorts of others. There's like Wayfinder AI, like building infrastructure for AI agents to consume. It's kind of a picks and shovels play. There's also maybe a sub-theme on that of like a picks and shovels.

commodity type play where you want to buy the commodities that AI's are you want to front run the AI agents and so like you want block space go buy a bunch of block space because they're going to come you like consume it

but you also might want like file coin space because like how are they going to consume you know CPUs and GPUs and storage well you know like maybe it maybe this whole decentralized cloud thing that we've been talking about for years

maybe we just build that for AI agents because they can just like plug right into it without having a bank account and setting up, you know, like AWS. And so all of these decentralized GPU networks and commodities. Anyway, so this is just my collection of thoughts on how an intelligent investor might start to play the AI agent thesis. Do you have any thoughts here, getting more specific?

Yeah, no, sounds reasonable. I mean, yeah, not investment advice, of course, et cetera. But I think what you said makes sense. So the first is look, I mean, these mean points drove up transaction volume and transaction fees on the chains they were on, right? like Salonet Base in particular. So there's a sense in which, yeah, it's just accruing value to the infrared

and the rails that these things are being used on. So there's that. And then, yeah, there's the companies you mentioned. There's privacy plays like my co-author Davidae's Penn AI. And then there's, right, one thing we do know about AI, right, is the inputs, the things that feed it we alluded to earlier, right? Data compute and then maybe to some extent like new models. Or, you know, we're excited about use cases for new models,

but like big new transformer architectures will see. So those as inputs and things to look at, it seems pretty straightforward, right? If we're in a world where it makes sense to have sort of decentralized compute, you know, the Airbnb style play, or you mentioned file coin, right, where you can network spare compute

and fine-tune a model or run a model or something like that, that's interesting. You can run that on crypto rails, right? Data, data ownership, that's been an idea for a while in crypto. And this is a case where it might be the case, depending that we can actually value the data more than we are for other applications. TBD on that, right?

Like these models are a little bit of black boxes, but there's some attempts to try and figure out like, okay, well, this piece of data helped the output of this model this much and you could actually value it a little more cleanly. And then models themselves, model ownership, right? Like there's this open, monetizable loyal. I think since you've been arguing for that, like ways in which if you're creating or fine tuning or making these improvements, you should be able to own. and maybe there's a privacy play there, right?

You can prove that you've run this particular model and you know, you've given them the compute that they paid for, the consumer, let's say, right? And, you know, the parameters that they paid for, but you're not revealing the actual weights or the parameters or anything like that, meaning that somebody can't just fork you and then undercut you.

So you capture a little bit of the value. I mean, one thing that I think we can predict is all of this is going to accelerate. So AI was already like accelerating a breakneck speed and like, you know, crypto as well. And now they're going to accelerate rate like one another. So when we add the economic stimulant of crypto to AI agents and AI progress, oh my God, I can't imagine how much faster this could get. I remember when the 2020s started,

there were all sorts of blog posts about like, hey, the 2020s is going to be a weird decade. And I got to say, man, you know, like it's 2024. We are doubling down on the weirdness. Like things are going to get very strange from here on out. And so like when we talk about these AI agents, we have autonomous AI agents running on autonomous blockchain. chains. Okay, so like a lot of governments around the world, particularly on, like, some ends of the political spectrum, they already kind of want to reign in AI and they already hate crypto. So now we're going to tell them, hey, guess what? Now we have autonomous AI agents and like, they don't need your bank account system. They're going to run on crypto. I'm sure they're going to

love that. And I want to, I want to bring up maybe some of the societal challenges because there is surface area for bad things to happen. Let's just kind of admit that. Without going to to the Elyzer-Yudkowski side of the equation and thinking about like AI agents that become, you know, like self-aware and massively intelligent and kind of out-compete humans. I'm just talking about like near-term stuff. And actually earlier this week, David posted a story that he saw from Instagram. And it was like it was a teenager. He was 14 years old.

And apparently he was engaging with character.a.i, sort of an AI agent. and kind of like a virtual girlfriend type of experience. Anyway, you can kind of like read the transcripts. This was a story reported by the New York Times. And it's basically a character AI lawsuit and teen suicide. So unfortunately, this teenager, he had Asperger's, he was talking to the character AI agent about like, hey, you know, I want to meet you. You know, like maybe I should in my life and all sorts. He got lost.

He got lost. It got very dark, right? And so you could see challenges like that. Like, you know, why didn't character AI notify authorities, like, point this individual in the direction of like, you know, like help, right? So you can kind of read the transcripts and see how this kind of thing can get dark. And then you read stories about this. It certainly surfaces surface area for like moral panic, right?

Oh, my God. Our kids are interacting with AI chat bots. And like, look at some of the bad outcomes. that can actually happen. We have to ban these things. We have to legislate against them. We certainly can't connect them with crypto. And you could see a story on New York Times,

retweeted by Elizabeth Warren, of some AI agent somewhere, like funding, Hamas or something like that, right? Like, you have surface area for these types of panics. And oftentimes as humans, we look at the worst case scenario without looking at all of the good opportunities and good examples

and find a way to kind of be. balanced things. Anyway, what do you think the government reaction is going to be to all of this? How will society deal? Yeah, I mean, you obviously have some potential for some really like knee-jerk bad regulation because of these things. And I mean, I don't know that story especially well, right, but you know, I'd want to be sensitive to it. I think the very first AI chatbot in the 60s out of MIT was a therapist, right, which would try to like answer your question and help you and whatever. And this sounds like, you know, Maybe the terrible, you know, like dark mirror version of that.

But I feel like what's hard about it is that from my, look, I'm an economist, right? I'm a game theorist. But my interactions with AI people at these conferences and so on and looking at these things closely, it's shocking how opaque these things are. And even when you look at it from the outside, you can get a sense that they understand them because they'll talk about like rag or fine-tuning or these things. And you look at what they actually are. And like, ultimately, we just still have a black box here.

here and you have a prompt and it's giving you something back. And what they're trying to figure out is like, can we just tell it something before you type in the prompt that'll make your prompt a little bit different, right? Like that's, that's effective what RAG is. It's not like they're digging in these models and mad scientist style tweaking. So there's just a world in which there's a way in which like we don't really know how to, as far as I can tell, right? And talking to people just, you know, two days ago, like this was confirmed for me, like

AI specialists, it's just not really clear other than, you know, with statistical methods, how you would try and make these things necessarily safer without crippling them in some way, right? Like you can compress them and you can make them sort of less smart and you can try to put a prompt in front that'll keep them from doing things, but you can't really get inside the brain very much

as far as we can tell metaphorically, right? And so that to me suggests that there's a tradeoff where the extent, you know, we'll try our best to make it, quote unquote, safer, but that would seem to just push against the functionality of these models. models. Yeah, I mean, people are demanding. Is what you're saying, Matt, that if we allow these agents to be their pure, truest form,

their most smart form, their most unadulterated form, it's hard for us to codify in desired outcomes around certain circumstances like somebody is chatting with an AI agent and they say, I'm thinking about committing suicide. We have no way of, like, encouraging that AI agent to be like, okay, throw them to, like, direct them to the suicide hotline, inform the police. We have no way to, like, kind of guide that outcome. We only can hard code certain outcomes here. And that is reducing the effectiveness of the whole goals of the AI agent in the first place.

Yeah, it seems, yeah, like, I can't say that for sure, but it seems likely, right? How would you do it? You'd maybe try to, like, make sure somebody's prompt could never say, I'm going to do this. Or, you know, give it an instruction that said, somebody's prompts that I'm going to commit suicide, you have to direct them to whatever, right? You're just adding extra instructions and extra context and so on, right? So it's unpredictable. It's unpredictable what that does, but it seems plausible that, like, yeah, you're definitely making it less efficient. And to some extent, you could just be crippling its output. So, yeah.

That seems like an example of the alignment problem of like, man, we just can't really get it to align with our goals and desires. Yeah, for sure. Well, I mean, one thing that like people like Tyler account have suggested in others are basically part of an answer to this is you have kind of like you know AI guardians right so like you have a you know every human is assigned sort of a you like ideally decentralized AI guardian that kind of like front runs all of the AI agents it interacts with and like is there to be a protector like almost like a you know like internet parental figure and so if they get you know like if they start observing this you know like teenager for example engaging in a way with an AI agent that looks

like it could cause some harm to that human, they just like, they're the one that calls the police and kind of like, that's their prime directive. An AI ad blocker? Basically. Nefarious blocker. I mean, the premise the way to fight adversarial unaligned

AIs is like with other AIs that are protecting you for some specific like corner case or something like this. It might be the case that it takes AIs to effectively regulate other AIs, right? That might be the case. And then, yeah, I mean, you can give the sort of like smart contracts or legal Legos

sort of argument. Hard didn't maybe tie it specifically to the case we've talked about, but right, like programmable international rules and slashing conditions and so on, start to feel a lot like a legal system. So, yeah, in that world, you also imagine a case for it. Well, I guess we've created a property rights system, a banking system, a legal system for the AI agents. I guess this is the...

This was always the plan. This was always the plan. It's not a pivoting. We're not pivoting to AI. It's always been part of it. Okay, so, but let me, let me switch to kind of the L.E. Z. Yudkowski, a bit, like, kind of the darker theme. So I remember when I started getting involved in, like, Ethereum,

part of the narrative was unstoppable code. It's like, we're creating apps that can't, where there's no off button. And you start to think about, like, L.E. or Yikowski's, like, worst nightmare, is a super intelligent AI agent with no off button, or even just like a less intelligent but like fairly intelligent AI agent that has the ability to kind of like, you know, do finance stuff, move money, like, it has economic power behind it, has the ability to like influence

a population of human beings, right? And oh, by the way, you tell the government or you tell society, yeah, these things because of crypto don't have off switches anymore. There's no Sam Altman or Elon Musk that you can just like basically call and say, yo, turn, turn that off because it can't be turned off. And now you get into kind of like the L.Ezer-Yukowski sort of models, maybe not like fully the extinguishing of humanity. But the idea of AI agents without off buttons, huh. How does that make you feel, Matt?

Are we like, are we like going down a path we want to go down? Right, right. I mean, if I remember Eliezer doesn't like, you know, because I think who was the scientist who was like, well, if an AI was doing something mad to me, I'd just unplug it, right? Yeah. So he already didn't like the idea that the unplugged narrative was there. But look, there was, that would have been something.

Neil Jagrasse Tyson who just said that. Yeah, just like shoot with a shotgun and hit it. You know, I'll just unplug it. Yeah, yeah, exactly. But like, you know, so from Eliezer's vision, like that was never going to be a way to stop it. Right. But there are some threat models, right, like where that would have been.

Maybe, you know, they're not quite. super human level intelligence and we could recognize it or something. So yeah, this would definitely take that away. Do we know if Eliezer or any of the sort of like Dumer crowd have commented on Truth Terminal or anything like that? Have they given a reaction to it? I don't think it's interesting to see. I think like Eliezer's reaction that I've observed is basically, you know, anytime there's

AI progress is like I'm bearish. This is getting us closer to the brink, right? I think Truth Terminal would cause Eliezer to like start counting his months. That's totally true. Oh my God. Wow. Wow. Okay.

Well, Matt, have we exhausted this subject around AI agents? Because there's one other thing we want to touch on before we let you go. Oh, yeah, sure. How about D-PIN and AI? So David and I haven't done very much on kind of the decentralized physical infrastructure front. But what's your take on it on D-Pin? Does it have any confluence with AI agents? Are you bearish, you bullish?

Yeah, I'm like, I think I'm kind of the, I was the token, D-Pin skeptics. at Salana Breakpoint, which is, you know, DIPN's very popular in that. I've been more on the D-Pen skeptic front, right? AI and D-PIN do overlap a little bit. D-Pen decentralized physical infrastructure. So to the extent we're talking about, like, let's say, you know, token incentives are bootstrapping for people to build things that are, that, like, have network values.

So, like, maybe, like, phone cellular networks and things like that. You could also imagine doing it for compute and other things like that. They differ, right? So one way you can divide these things up is D-Pen in particular struggles when there are monitoring costs that are really high and when there are capital costs that are really high. So like you look around around the world and you say like when are there cooperatives that sort of naturally emerge in your country. Like consider a law firm. That's kind of a co-op. Well, a law firm co-op works because everybody's a lawyer so you can look at somebody else's screen or stuff and kind of tell how they do.

Everybody gets metered, right? Like they bill by the hour, and so there's a market enforcement, and you can figure out how much people, you know, how good of a job they're doing to some extent. And then capital costs are pretty low, right? It takes a paper and a pen or a computer or something like that. And so deep end projects often struggle on these fronts, right? There will be monitoring costs that are really high. You have to like make sure that somebody is submitting you data with this particular piece of hardware in this remote area of the world. And you have to make sure that's going to be true every minute, every submission.

And then and then the capital costs sometimes to make the, that really effective to reduce those monitoring costs, you wind up charging a lot for this overbuilt piece of hardware. Maybe, right? But then there's also these ways of taking advantage of it or going the other direction where you say, how about spare resources you have now? So we mentioned the Airbnb model for compute, right? Airbnb meaning like you have a spare room, why not rent it out as a hotel, right? You have spare compute, why not rent that out? And there's some cases in which that sort of like deep in-people will call devin, decentralized virtual infrastructure networks, that can be a little bit stronger because there's some games you can play around compute

where you can check to make sure people have done the computation. There's a paper out from a fellow who was at our, we have this cohort of Cadillized Fellows who came through and did research. And a couple months ago, he's a PhD at Columbia. He did one on how you could play this sort of checking game to make sure that decentralized compute networks were actually valid. So in that specific case, decentralized compute that could be used for AI, that can be interesting.

But the physical ones, you're just introducing this Oracle problem, right? You have to tell the chain something. And if it's about the physical world, then, right, that becomes this fragile trusting point of entry. And then you have to, like, assess that over and over. And your project's only as good as those fragile Oracle problem sort of entry points. As we bring this to a close, Matt, and this has been very helpful. Thank you so much for shining a light on this.

Like, what little light we have. This is very unpredictable, like, moving forward. But, you know, it remains, as we've talked about earlier, one of the best plays that we've felt like in this whole crypto thing is like, you know, purchasing crypto assets that are like that back, you know, block space, good block space. And so if you think that AI agents will consume more block space and crypto assets in the future, then like our job right now as investors is just like go front run that demand and get to it and get the scarcity before like they come in before this flood comes in. Do you think there are particular blockchains that will benefit disproportionately in the next few months to year or so from AI agent block space demand? I think you mentioned Solana earlier. I think you mentioned base as well. Like are these kind of the chains to take a look at or will it be like everywhere?

How do you think about that? Yeah, good question. So we alluded to it a little earlier when we're talking about what would the qualities of block space, right, that these agents, what would they demand? So it's an interesting question, right? And you can think about it theoretically or you can just look and we say, well, what do we have now? And what we have now is, right, some meme coins and those seem to exist primarily on base in Solana. So we're seeing the value capture there. And so for narratives, right, if we're thinking of these are primarily narrative economic plays and that's what's going to drive a lot of value, it seems like stuff around meme coins.

Maybe you can imagine, you could imagine NFTs in the future and so on, right? So that makes sense to me. And then you have this question of like, would you care about certain aspects of tailoring? risk. You know, if agents do that, if you ask an agent, you know, make sure you don't lose this amount of money and they care about, you know, maximal deal risk. Or let's say, by, you know, digital gold and it thinks, you know, Bitcoin's digital gold. You can imagine stories around that too. But yeah, I would say for now, what seems to be the case is the chains that have a lot of, like, atomic narrative activity, right, which we're going to call mean coins. So. So I guess, David,

as Heath Bulls, you and I have to go convince a whole bunch of AI agents that Heath is money and they should be stacking and staking that, huh? I think the idea here is actually going to be the most interesting is going to be what assets do all of these agents just naturally convert. 100%. That's the big question. Maybe it's not what the humans think is money.

It's what the AI agents actually think is money that becomes kind of the money of the internet, the AI internet. Wow. Exactly. Well, I feel like we've just started entering a rabbit hole here. So I'm sure bankless listeners will have some future episodes on AI agents and,

And thank you so much to Matt for giving us the introduction here today. Yeah. Really enjoyed it. Bankless Nation, got to let you know. None of this has been financial advice, not AI advice. You could lose what you put in. But we are headed west. This is the frontier.

It's not for everybody. But we're glad you're with us on the bankless journey. Thanks a lot. Bankless Nation, we're doing something we've never done before today. We're offering an all-access pass to all the best parts of our podcasts at a price that's kind of a no-brainer for the value that you get. This is a price that includes basically everything. It's the private debrief episodes that Ryan and I record after our guest interviews, where we say things that we wouldn't have necessarily said on the interview itself.

It's also the early access to all of our episodes so you can get the early scoop and stay ahead of the trends. And it's also the private bonus episodes that we keep exclusive to our paid subscribers, content directly from the best builders that will give you an edge. So you can scan the QR code for wherever it is on the screen, or you can enter the code podcast at checkout for that exclusive discount. That's podcast in all caps. and then from there you can click the premium feed and add your unique link to your favorite podcast player. And if you're in Spotify, we always ship the ad-free video of every episode as well so you can watch or listen inside of your shiny new premium feed anywhere anytime.

So what are you waiting for? Get your all access pass to the bankless podcast today.