Epicenter - Learn about Crypto, Blockchain, Ethereum, Bitcoin and Distributed Technologies - Emin Gün Sirer: Avalanche – The Future of Crypto: From Gaming to LLM-Powered Smart Contracts

Episode Date: May 11, 2023

As the second Avalanche summit drew to a close, Sebastien Couture and Sunny Aggarwal sat down with Emin Gün Sirer for a fascinating conversation exploring future innovation narratives in crypto. Longtime friend of Epicenter, Ava Labs' founder and CEO shared his unique insights on the industry, from both a founder, as well as a former academia perspective.Given Avalanche's consensus & architecture, it boasts high TPS and scalability properties, incentivising founders to build better products in the AVAX ecosystem.One of the most controversial narratives is that of LLMs (i.e. ChatGPT) powering smart contract development by bypassing programming languages and even bytecode altogether. The question arises if replacing the human third-party with an AI fixes trust issues or not, but only time will tell which piece of code we can actually consider 'law'.Topics covered in this episode:Emin’s background, from academia to running a multi-billion dollar companyMain takeaways from the 2nd Avalanche summitInnovation narratives in the Avalanche ecosystem. NFTs & blockchain gamingThe Avalanche architecture, consensus & scalabilityAvalanche’s evolutionChatGPT smart contracts & no-code blockchainsAuditing AI & securityInnovation vs. StagnationBitcoin, BTC.b & ordinalsBridging Cosmos with Avalanche: IBC x WarpBest & worst products built by cryptoEpisode links: Emin Gün Sirer on TwitterAvalanche on TwitterAva Labs on TwitterShrapnel on TwitterGunzilla Games on TwitterThis episode is hosted by Sebastien Couture & Sunny Aggarwal. Show notes and listening options: epicenter.tv/495

Welcome to Epicenter, the show which talks about the technologies, projects, and people driving decentralization and the blockchain revolution. I'm here in Barcelona at Avalanche Summit, 2023, and Sonny and I got to sit down with Iman Gonsir, CEO of Avalavs and founder of Avalanche to talk about this event, the Avalanche ecosystem, and also some exciting new ideas that are being explored in the Avalanche ecosystem. about bringing GBT to blockchains. So this is an hour-long conversation where we went deep and really got to pick Immien's brain about where he sees blockchains going in the future with regards to artificial intelligence

and large language models. So I hope we'll enjoy it. So without a further delay, here's our conversation with Amin. Amin, thanks for joining us. All right, Sebastian. Thanks for being on the show. once again.

So this is going out on Vienerop, but also on Epicenter. And so our epicenter guest will remember you from your numerous appearances. Last time was about a year and a half ago. And you came on to talk about avalanche. And before that, you were on the show to talk about a bunch of stuff. And you were all your research previously from when you were at Cornell, Bitcoin NG, Selfish Mining. You could also print the paper with Vlad about the Dow,

which turned out to be mostly correct. And so, Mattoon. Yeah, so, you know, for those who don't know you, maybe get a background on your trajectory and how you got to become the CEO of a major blockchain. Sure.

So let's see, I was a professor at Cornell for many years. I started out there in 2001 at the time when peer-to-peer systems were just taking off. And I was always fascinated by self-organizing systems. I was always into building really complex. like systems that could give you a strong guarantee and stand behind it. That was sort of my driving motivation across all of my research. And I started out by working on a peer-to-peer

currency called Karma for file sharing networks. It had proof-of-work minting in it in 2002. As before, the Bitcoin White's paper. It did not have consensus via proof of work. So Satoshi had a huge step up from the work I did. But I was thinking along those lines a long time ago. And in any case, I've been working on everything blockchain for a long time now. Worked on characterizing the security of Bitcoin. We found the biggest known flaw in Bitcoin, known as selfish mining. We worked on doing a bunch of other things, on characterizing decentralization in blockchain,

on securing coins at rest, on building Bitcoin NG, this protocol, that is the next generation. That's faster, better. And then most recently, I've been working on lightweight consensus protocols. And that work led to the avalanche system. And I spun that out and left Cornell about a year ago to focus on it full time. What has been the biggest change from going from being like an academic to running a multibillion dollar project? That's a good question.

So I think about that all the time. So we were at the forefront of research and development. I was worried before I left academia that I would not be able to do research. that it would be hard to attract top-notch talent, that we would essentially just kind of fall into a rut where you try to digit to monetize something. And that has not been the case. I was so pleasantly surprised that we are still actively engaged

in cutting-edge research in systems. And so that's been good. And that was surprising to me. The other changes in my lifestyle are that it's insane how my life has changed from doing whatever I used to do to doing whatever I used to do plus so much more now because I have to do these other things as well that relate to running a large company.

So that's been the biggest change for me. But overall, it's been a fantastic change. So in academia, you always have to wait for someone else to take your ideas and take them to completion. In the process, good bits are dropped. Or maybe they take a good idea, but they execute poorly on it.

And when you are in command of your own destiny, it's a very satisfying feeling. So it's been a great change for me. What is the part that you miss the most? I miss teaching. I miss teaching so badly. I have dreams where I teach. And I used to enjoy it immensely.

And it's actually very addictive. And I used to love that process. And it's also very satisfying because you can see, like literally in people's eyes when you've introduced a concept, like multi-threading. They don't really know how it works until someone explains it to them. And those aha moments are so, so satisfying. So I miss that a lot. I do not miss the class size. So towards the end of my career, I started out teaching classes that were like 40, 40 students.

And towards the end of my career, I was teaching 480. Wow. And this computer science area has exploded. Well, 480 is a lot, you know, just I had like 30 TAs. Just the TA group alone was similar to play initial class sizes. That sounds crazy. I mean, anybody who's interested in, you know, going back and, you know, learning more about Emmyn's previous research

should definitely go back and listen to those old episodes of Epicenter. They're old, but they're still super relevant, and I think. And they're fun. Yeah, and they're fun. Yeah. I'd go back. Lots of drama back then over a technical issue.

Different drama. Different drama. Yeah, we're here at this, like, beautiful venue. And this is the second Outalunch Summit. It's honestly, it's like very impressive. conference and and so wondering what are the main takeaways for you after three days here? For me, let's see. I think this is a occasion where we come together. And you know, post-pandemic, we all work in our little universes, so I kind of get cooped up and it's easy

to lose track of where we are. It's easy to lose track of one's community. Or, you know, you always have some sense of who they are, but it's a nebulous thing you construct it in your head. And this is the opportunity to see them. So my main takeaway is just the vibe right now in the middle of a bear market is so much more positive and so much more growth oriented than I thought. And so I had to do a huge adjustment in my head from where I thought we were to what I actually see. So I saw a lot of companies. They're really interested in building blockchains and something that will resonate with this audience. They want to create a custom-specific you know like an application specific chain so there's quite a few of those around vCs there are

far fewer VCs they're very very loud online right and these are people are like oh yeah like they act like they are fully committed but when you look around they're not actually around and so they've been most of them have been wiped out and or they're not they don't seem to be as committed in actual fact as they seem to be making noise online so i wasn't actually at last year's avalanche summit but from everyone I've talked to here, they said that they enjoyed this year as much better because last year they felt there was, yeah, a lot of suits, a lot of VCs,

but everyone out like this year... Far fewer. Yeah. Was this an impact of just like the market or did you guys make this like an active effort of how to? No, no. No, no, we invited all the VCs as always.

And, and, but I think in general, like I was just at consensus and I saw far fewer VCs there as well. I think overall as a sector, this area is not attracting as many. dollars as it used to they're probably back home there's still tons of funds like there's more they're sloshing around they're sloshing around uh they're probably back home looking at uh i i deals but that's where the money is going um but uh but i i love this this is this is the vibe that i

really wanted to cultivate just let's just build stuff there's so much undelivered promise so far and uh you know all this like coin coin centric vc investments where people create coins just for the sake of creating them yeah that wasn't healthy at all and i'm glad we're sort of flushing that out of of our stuff. So what's the most interesting stuff you've seen this year and, you know, I mean, actually, maybe like, like, forth answer that because if I saw something out, then I'll get the mail from the people I left out. Okay, well, then maybe at a higher level, like, what's the kinds of innovations, the sort of more general industry directions that are, that are sort of playing out. And specifically, like, I wasn't here last year. And, like, to be honest, like, I've discovered

a lot of the, I've lost ecosystem here, because I spent a lot of my time in Cosmos, like, sending And so how has the avalanche ecosystems have evolved in the year and things that people are working on? Yeah, great question. We have many more sub-ness deployed this year than last year. I think last year we were trying to get the subnet idea out to the masses. This year, we've done that.

We are already seeing a lot of traction. There are more than 55 production sub-ness, even as we speak. There are hundreds in testing, and there are lots of great sub-ness with great chains in the, pipeline. So one of the big trends I saw this year is gaming. They love having their own chain. They love having their in-game economy on their own blockchain, on their terms with their own virtual machines. So we're seeing that happen. We're seeing that happen with AAA games. So for example, Shrapnel is here and they have a great game coming out. It's like CounterStrike Plus

Plus. I don't know if you guys got a chance to play with it. Those of you who were at Consensus, they had a great booth there. I played it there. Just a amazing game, far better than campus. And the game dynamics are fantastic. The artwork is fantastic. We have another game. Poor gamer yourself. I am not. I'm a terrible gamer.

You know, the last game I was good at is Frogger. Okay, so let me date myself. You know, I was there when, like, Pong was a big, you know? So, not much of a gamer, but these games, like, they really, I don't know, it's just they're so realistic, so beautifully done. There's another one, there's another AAA game called Gonzilla. That's coming out soon later this year. Amazing, amazing production values.

Like the universe, I played that game a little bit. They flew me out. They 3D recorded me, kind of like they did to Gullam or whatever. And they made an NPC out of me in the gate. So there's a copy of me in the game who looks into the camera and says dramatic things, and you can unload a clip into my head, and I'll respond and say dramatic things again. It's really fun. And I loved it.

And if you play these games, you just kind of, lose yourself in the game. Like the, you know, you can replace your arms. You can replace your legs. There's like mecks and stuff, bots and drills flying around. Just CounterStrike could be so much cooler. And these guys are delivering on that promise. So, and then, of course, everything you buy is an NFT in the game.

Amazing, amazing show. How do you, so, you know, I've definitely noticed a lot of focus on gaming here. A lot of people have been talking about that. It feels like right now there's just like weird culture war, between the much wider, bigger gaming community out there and crypto. My take is I think they're still mad at us for stealing all their GPUs. But how do you guys plan, like, you know, like Discord added crypto features and then they got so much backlash.

Yeah, that does undo it. How do you guys plan on like, so I see all these games that are being developed right now, but how do you plan on reaching out to this like more active gamer community? I think, so this is not my domain at all. We just provide the platform for people to thrive on, but I did play with the games, and I know what the game designers did. It is so smooth that I think what people really react to

is the crypto community's propensity to push coins into everything and to put cumbersome sort of front ends in front of stuff that you want to do naturally. So in these games, when you get a gun, that's an NFD you just received. And it's not part of, like, it's not in your face. you didn't have to buy coins, et cetera.

So it's so natural, and these people end up interacting with a blockchain without knowing they're interacting with the blockchain. I've said this many times before, that's one of the signs of success that people are using your system and they don't even know. So if we're not in people's faces, we're not going to get the backlash. It's that's so.

Cool. Let's dive into the nitty gritty a little bit more in some of the technical aspects of Avalanche. And you made an announcement yesterday that we kind of titled this episode as GPT for smart contracts. So we'll also talk about that. But let's maybe just start with a refresher on how Avalanche is architectured and the different parts of that, like the different chains of the C chain, the P chain, the X chain, so that people get a sense. So like how that all works. And then we can get into some of the more in-depth that thing.

Sure. So there are a couple of things that make Avalanche unique and interesting. The main one that got us started down this path of issuing a new chain is the new consensus protocol. It's unique. There is none other like it. And it achieves consensus via this mechanism that's very similar to gossip for those people who are familiar with it, but it's a consensus mechanism as opposed to just the gossip information dissemination mechanism. So it achieves agreement at the end of the day using this very lightweight mechanism

that allows our system, which you consensus at about a second. We advertise a second. We're really, actually, in actual fact, you usually below a second to finality. And it can do so

without compromising decentralization at all. We can have millions of nodes participating in every decision and still achieve those amazing turnaround times. So that's one. The second thing that's interesting about Avalanche, like Cosmos, is

its ability to support more than a single chain. So Avalanche, in fact, there are three of us. There aren't that many chains that have this architecture. So prior to us, initially there was Bitcoin, one asset, one chain. Then came Ethereum and other chains like that, which are multi-asset single chain. And now you're seeing the rise of the next generation, the highest tech in the space, which is multi-asset, multi-chain. And in Avalanche, you can have multiple chains in parallel, each running their own virtual machine, each with their own gas asset, each with their own staking asset. And that allows people to build a gaming chain, for example, that's completely shielded from the load on the other chains.

We have a certain chain, and we have a generalized smart contract chain called C chain. So it's an EVM, and so you interact with it the same way you do with solid deep programming on it. It's a much faster version of Ethereum, essentially. You can create more C2, C3 chains if you want to. We have this other chain called X chain for assets. So if you just want to easily issue a token, you can do it there. And then we have this coordination chain called P chain. Those three are the ones we started out with. The coordination chains for finding the other chains that are in the system. And as I mentioned, there are about 55 of them even as I speak. So that's our big architecture. It allows anybody to create their own chain subject to their own rules with their own validators. We don't have

centralized components. You don't have to go through a hub to communicate. We have a And so IBC-like communication mechanism that we call warp. So any blockchain can talk to any other blockchain. And overall, so those are two things that make us unique. The third thing I think is a mentality thing that is very important to us, which is to use the best of science in everything we do. So it's not just the platform itself, but our bridges use different technology. We've been building exchanges that are using secure technologies such that even the exchange operator cannot misbehave.

So there's a lot of tooling that's happening around Avalanche that's essentially bringing new science into the space. So I want to talk a little bit about the interoperability protocol, the warp protocol. I think for people who listen to this channel, there's an understanding of how IBC works with light clients that are able to, to verify the state of another chain and essentially mint assets on another chain and sort of burn assets on the initial chain. And that's how we arrive at being able to transfer tokens from one chain to another. In Avalanche, it's totally different because there is no light client by virtue of the consensus mechanism, or at least that's how Sonny's explained to me earlier today.

Yeah, he's not drawn. You can have light price, but it doesn't work. Correct. Warp does not work the way ABC wins. We use aggregated signatures across the validators that's comprise a blockchain. So if your chain wants to send a message to me, essentially what we do is we construct a quorum signature from the set of nodes, the validators that make up your chain that say, in this chain, at this moment in time, such and such an event happened. Perhaps you're sending

a message to me or perhaps you burned coins, what have you. And so that becomes a message in a well-understood parsable format by any other blockchain that wants to consume it. And then that can be taken by any other chain and used to do the corresponding action, an invocation or creation, minting of V-cois on that other chain. How scalable is that? Because one of the nice things about Avalanche Consensus Protocols that can scale to millions of nodes, but these aggregate signatures, how well do they scale? That's a great question.

And so there are lots of different cryptographic techniques. We're compressing those signatures. The technique that we use comes from, you know, from BLS signatures. So Dan Bonnet and his team at Stanford, it's incredibly scalable. And at the time I was worried about this because, you know, there is cryptographic constructs. And I think the Ethereum community knows very well that a lot of them don't pan out in practice. There's some feasibility issues that you run into that are not evident in theory. So BLS signatures, the way they're implemented, is incredibly fast.

So a warp message can be constructed in hundreds of milliseconds across. thousands of validators. So if your blockchain is, you know, again, say 10,000 validators or less, you're looking at less than a second for messaging. You're typically looking at about 300 milliseconds for message construction. Still needs to be parsed on the other side, but that's super fast. And you can become an input to a smart contract on the other side. Yeah, I guess that's a nice thing about BLS and kind of aggregate as you gossip, right? You don't have on like other Android signatures where you need N-square communication. Right, exactly.

So what are some of the ways that the Avaland Consensus Protocol has evolved over the years? I know, like, there are some design goals that got set out with, making this very leaderless style protocol, and then those have sort of had to shift over time. Yep. Can you talk about how it's evolved? Sure. So there are, so the Avalanche Protocol itself, when it came out,

it's actually it's not a single protocol there's many protocols the new family entirely so for those of you that are listening that are familiar with protocols like Cosmos

and let's actually take Cosmos is great because a very well understood sound of really really nice protocol we call that it's in the classical family there are other protocols like ETH2, EOS, etc

they are signature aggregation protocols they don't have a view change, as you know, processed and up. I don't know how technical we want to get into this, but... Well, okay. So there are a bunch of protocols that are, you know, fairly specific. They do signature aggregation, which becomes very, very cumbersome, the more validators you have. Avalanche, when it came out, came out as a huge series of protocols. There were like four or five in the initial paper, and it's grown since then. And I think there are two different, two surprising things that happened, when we came out with the initial avalanche protocol, it was leaderless. And it's still,

that protocol is still there in every way. All of that code is still there. So in any slot, anyone can propose. But if you do run a protocol with that sort of, that sort of modality, then you can get into situations where there's a lot of conflicts because, you know, we've all been at conversations where two people are constantly talking at the same time. There's a pause, two people start the sentence at the same time. Then they have to stop, back off, look at each other in the eye, you know, and then you start again, and sometimes it can repeat itself. So to stem that, we added a soft leader mechanism to Avalanche.

So it's not a hard leader. It's different from classical protocols. There's no view change. There's no leader election. It's just, you know, like in this slot, you know, we're going to give. We will all pause a little bit more, but Sunny doesn't have to pause as much because it's his turn. So that's the kind of heuristic that we added to cut down on congestion in the network. It was incredibly effective.

But the sort of the fundamental nature of the protocol underneath did not change. But more recently, we made another change. It's just maybe 10 days ago. We did the Cortina update. And so, you know, one of the products, when we went out, we had two different consensus protocols that we used and two different data structures that we built. The C chain was an EVM chain with an, you know, Ethereum virtual machine running on it, that created a totally ordered linear chain.

The X chain was an actual DAG protocol. So we created a graph, a directed acyclic graph. And, you know, as an academic, you know, it's incredibly appealing. I was like, oh, yeah, a DAG is so much better than a chain because you can grow it, you know, it's you can work on this side of the graph while you and I work on this other side of the graph and it can proceed in parallel. So theoretically, it has far more capacity. So then we went out with it. So one problem, of course, is it's more complex. That's not a big deal. We have so the best developers and, you know, we can handle complexity. But the second thing that I discovered,

to my surprise, is that all of the exchanges in the world, are written for linear chains. So even if you have this DAG, they don't know what to do with it. And they keep wanting to ask a question like, did this happen before this or vice versa? And in a DAG, you have called current operations. That was the appeal of the DAG.

So I spent countless hours talking to Chinese exchanges and saying, look, you know, the question you're asking is malformed like that. Don't be asking that question so we can go fast. But then they say, no, our systems are designed this way we have to ask and smash it. So that ended up forcing us into rethinking the DAG issue. And with the Cortina update, we're like, okay, look, you know what we're going to do?

We're going to take a hit on performance on the X chain. Not that big deal. We still have ample performance. So we'll get rid of the DAG and we'll build a linear chain just to make exchange integration simpler. Also, not just that, but also to make warp integration simple. So now Warp works across all of the subnets. and we got rid of the DAG, which I think was, you know, a deep to hell I miss it as an academic,

but as a practical person that wants to stop to work, it was one of the best rules we did. Yeah. When Avalanche becomes, you know, the biggest chain that we'll teach them about that. Exactly. Exactly. At some point, the world was not quite ready for this.

This mechanism you talk about with the, sort of it's like, let's all kind of soft agree that someone is going to come first, and, you know, we'll all respect that. How does that respect breakdown when you have these MEV competitions start to form? It works just fine. So if an MEVer jumps in, that's fine. We're going to prioritize the person who's in the allotted slot.

So that does not give any advantage to MEVers. The way to win the MEV game on Avalanche is actually very straightforward. You just need a crap ton of validators. And to do that, you need to stake a crap ton of AVVACs. So in a sense, we turn the MEP game on its head and say, look, if you want to play this game, it's a game that cannot be stopped. People will want to play. We want to game your protocol no matter what.

It's like there's a well-understood way. You need as many slots as you can to your name, and the way to acquire that is to stake a lot of AVAX. And so that's exactly how that works. Let's talk about this AI stuff. Yeah. It's crazy chat, GBT, smart contract writing, which is not actually like when I was listening to the talk or someone described it to me earlier

it was like oh we're going to use GPT to write smart contracts that's not at all the idea I think the idea goes even further it sort of abstracts the way the entire like oh yeah sort of way that we think about writing applications or giving instruction to like a machine so yeah talk about this this idea and yeah well what's what's what's with it what's gladly gladly so it's a very obvious idea so once I tell everyone what it is they'll be like yeah I see how that could work and and it's also a very bimodal idea it's either a great idea or a terrible idea it's not an in-between it's not incremental it's just not and it's also not something I heard before but once I say it'll be obvious so here's the idea we are starting I'm super

excited about this we're launching a new project to build a new sub-euvre on the avalanche that's powered by AI, specifically by large language models that are using these generative transformers, a pre-trained transformers known as GPTs. So what are these things, right? So these things are, you know, at the end of the day, they're giant matrices. But they really are engines that have been trained on the totality of digital stuff we have on the internet. They've read everything on the internet in every language. And they've constructed this intelligence, so to speak, that knows how to react to situations that are provided to it in natural language. And so what's the proposal?

We build a new blockchain. This blockchain has, as its execution engine, one of these trained language models. It has a generative pre-trained transformer. So this GPT is in there. Think of it as a no-code blockchain, a blockchain where there is no transaction format. There is no bytecode. There is no solidity. There is no wazom.

There is no other programming language whatsoever. When you want to do something, you write it. In what language? In whatever language this thing is trained, and it happens to be almost every language. You write it in English. I give to Sebastian five AVVACs. I write this in text as my transaction.

With those letters, it's like asky characters. And that's part of the blockchain. And then you can write something that's kind of like a check. But more expressive, you can write something like, hey, I give to Sunny, you know, 500 AVAX, assuming that he can complete his fundraise for his new movie that he's making. If not, if he can't do that, I want my money back. That's not something I could write on the check that my bank gives me, right?

And it's being executed, whatever you write, is being executed by a very smart, neutral counterparty. That is the intelligence inside that AI engine. So there's no more programming. We all become programmers. We can all write smart contracts. You can start something like, I now start a sequence of interactions using the rules of chess. And Magnus plays the white pieces and, you know, we play the black pieces.

You know, I don't have to define the rules of chess. Yeah, right? If I were writing a program, it would be this long, and then it would be complicated and then I have to do all that, like, checkmate, checking and so on, it's expensive. So here, this is just like, I rely on the inherent intelligence that this thing built. I can say this contract will divide the, you know, we wrote a song together, let's say, we'll divide the royalties between three of us fairly.

And we now appeal to the notion of fairness built into the AI. So it's just, it's an amazing new future. Is it crazy? Could be. Could it be a bad idea? Yeah, could be. It was an experiment worth doing. So in the talk you described this, this notion of interpretation.

So we have like an intent. And so we vocalize the intent as like language or like, like you just said, right? I want to send these tokens to this person under this condition or like, well, that's, that's why we describe things as humans. And then we translate that into into code, right? Like rust or go or solidity or whatever. And that gets translated into bite code. And the issue is that between these different intents, well, there can be.

So things that are lost in translation and worse yet, like bugs, right? So like reentency attacks and things like that where you say, okay, like I want to do that. But I haven't thought that there's like reentency or there's this particular bug and this interpreter. Does this idea imply that in the background there's actually code being written? Or are we sort of thinking about a different type of VM here that's like? Yes, we are. Okay.

So it's like language models a bytecode or is it language? No, absolutely no code anywhere. There's going to be no byte. code. There's no code construction. The AI is not creating code for you. The AI is interpreting the actual text you write. But at some point, at some point, like, there has to be a transaction, right? And so that transaction gets run by a validator, and that validator is executing some machine code. No, it's executing GPT. It's running the matrices that actually power the AI. So you're having a conversation the way you have it with chat GPT. And the validator is a, and the validator is

executing whatever the language model is coming up with in response to the prompts that are coming in as transactions. So there's no code construction, absolutely no rust, no bytecode, no like bit with language. But the state change, like, how does the state change happen? Yeah. So that, I guess my question is like, at what point does the GPT model interact with a state change? The GPT model, every validator has a GPT model in it, and it has an account, some kind of an account abstraction where it associates token balaces with addresses. So if in response, if you tell it to modify one of those things and it's pre-prompts, you know, in the Genesis vertex, so to speak, instructed to, then it affects that change. So if I say, hey, you know, I would like to give

five A-backs and I am authorized to do this, then it's going to make that change in its central account obstruction. So we have one AI and one accounting ledger, so to speak, that's kept in every validator. The AI engine has to be deterministic, by the way, so that's one of the restrictions that this approach required. And so when the transaction is seen at the validators, they all see what's being asked. They apply the validity rules that they have been pre-prompted, and then they make the effect that change that's required by that transaction. How is security going to work in this world? Because one thing I've like, you know, chat GPT is great.

But like as you use it more, you start to realize it does some weird things. Yes. Absolutely. You know, hallucinates. All right. And then you have got like, amax. Also, like, if I do like trans-if I like, even when I'm coding, like, you realize when I, you know, it's great for generating test cases and all this stuff. Then you start reading the code.

You're like, oh, that's, you're making some bug. You should have to go review. of everything. And so if there's no standard, like today with like smart contracts, we can like that, if we have bytecode, we can like, you know, do actual like, you know, if they're all full, you go full formal verification, at the very least you have some way of like auditing that this thing is doing things correctly. Yeah. While with the GPT system there's no rules to audit. Absolutely. So let's tease that the part there are so many concerns. As I said, this is either a great idea or a terrible idea and nothing in between. So let's see. One of the big criticisms with these

machines is these new emerging machines is that you don't know what they have learned, right? You can't audit what's inside, et cetera, et cetera. I fully agree with that. We have to just adopt it. We have to own it. So your Tesla drives itself. Is it going to go crash every now and then, you know, until it's trained to be, to not do that, it will. It's kind of like a learning, you know, 15-year-old kid who's learning to drive. So it's same with this. So we will have to make sure that whatever AI we use is actually intelligent. It does actually have a good grasp of what it's supposed to do. So I'm hopeful on this front though that we interact with each other. I don't really know what's inside your brain and but somehow we managed to all get along all like, you know, one of the

eight billion of us. So, and I'm very hopeful that the AI evolution will actually speed up, and we'll build these engines that are really, really smart very soon. It will happen far faster than people think it will. So that's one thing. The other thing, prompt engineering, hallucinations, etc. That's a huge concern. This is why I want to run this as an open project. This is why I need everyone's help because we need to experiment with this.

So undoubtedly, if I were to specify a lending service in English right now, I can do it, right? There's a lending service. It has a bunch of pools. The pools do this and that. And, you know, on the back end, the machine will keep track of accounts and do the lending. But undoubtedly, my specification will leave something open. And this white person like, you will come along and you'll be able to say, you know, whatever.

Redefine what, you know, Ging just said to, you know, change all the commas into whatever. You can do all sorts of funky things to trip it up. And what I think will happen over time is we will come up. with libraries of preambles so as to restrict what comes afterwards. And that, I think, will give us a modicum of safety. But it is a multi-year project. I don't think like this is going to be a fun project. I think it's fun.

So the lawyers are the only ones who are going to keep that. Or we all become lawyers, right? So I agree with that. Lawyers who know how to write fine print. You know, the fine print is there for a reason. And so it'll turn into what we call this thing COA, coin-operated agents. She's going to turn into preamble for COAs. I mean, this kind of extends, this idea can work on a blockchain,

and certainly it would be cool to see that come into existence. But it could also extend to any computing environment. You know, we could have like a JavaScript VM in our browsers that executes code in a similar way. So are there other attempts to do this sort of thing outside of crypto? Is it something that was inspired by other? It's like a very novel idea that I think just expands outwards. I'm glad you think it's novel.

No, I was just kind of sitting at home thinking about AI stuff and thought, hey, we have to put this into a validator. I haven't seen, I mean, everybody and their brother is trying to work on the next GPT engine. And these things are going to find their way into all sorts. self-tooling. But, you know, undoubtedly that's happening. But I haven't seen anybody

talk about putting it in a validate into blockchain a phone. Interesting. Yeah. Yeah, there was just like interesting article. I was just talking about before the episode, but like a memo that came out like from a leaked, leaked memo from a Google employee talking about like,

you know, they're competing with open AI, but he's like, actually the biggest competitor right now is that the open source LLM is like, so Facebook released their open source. element. Like the model is much smaller than what like, uh, GBT through like open AI and Google have. But it's actually from a, you know, UX perspective is out like just the open source iteration on it is now competing. And like, you know, in surveys, they're like, oh, yeah, these are over 50% of the time people like don't care between these open source ones versus like chat. So I think building this like open platform like a neutral platform in a public way is very important.

I'm really excited. I'm happy that you're like really thinking about this. Because one of the things I've been a little bit, like, admittedly, thinking about it like for a couple weeks now is like, is crypto like a great stagnation technology? I hope not. I don't think we're stagnating at all. I know we innovating quite a bit.

So my question is like, I feel we were in a period of great stagnation. Yes, that's true. And that's why I'm here, by the way. You know, I was sick and tired of listening to crypto Twitter mull over the same damn topics over and over. How do you switch to proof of stick? Why does it take a six years to do this? I don't know. How do you handle identity management of the blockchain?

I don't want to hear anybody else talk about this. That's why we have this other push on institutional blockchains where I just want to show people, by example, this is how you do this. It's not rocket science. It's super pragmatic. At the avalanchevary, very pragmatic. And this is how you do it. Let's put an end to this discussion once in four. all, we don't need 15 working groups and 50,000 white papers and so on around this.

Very doable. Very pragmatic solutions exist. I'm telling me, I mean some of a little bit of metal layer, which I felt like crypto is very much about, like, taking the state of the world and, like, redistributing power. Yes. But it just feels that, like, wait, we're entering this, like, inflection point on, like, technology, like, power dynamic, like, trying to redistribute power doesn't make sense versus, like, the real power distribution.

Like, you know the story of the dreadnought? No. I know dreadnought. Yeah. So I believe dreadnoughts caused World War I. So what the story of the dreadnought? So in the 19th century, Britain controlled the seas. Like, you know, 10 times as many ships as the rest of Europe combined. At which you control the seas, you control the world. Then one admiral, his name is Jackie Fisher, he commissioned the creation of the dreadnought.

It was the best ship in the world. It was faster, more guns, further guns, better defenses, everything. And what it did was it made every other gun in the ship in the world obsolete. And the Germans saw them, I'm like, oh, we want one too. And then the British made one and then the Germans made one. One of the kind of did this like arms race. But you went from a world where Britain had 10 times as many ships as the rest of the world to a world where Britain and Germany had the same number of dreadnoughts.

And so it's like technology in like just like, you know, creative destruction is actually like when your past is how like real power redistribution happens. And so I think like crypto is that's what I meant by stagnation-based technology. It's not innovating new innovation, if that makes sense. Maybe that's controversial. I disagree with that. I disagree with that. There's so much energy like just behind this, right? And and I think we are addressing all of those things that, I think you're right, that for a long time we were kind of, you know, whirling around the same drain pipe without much progress. But We are solving the custodian solutions. Like, the institutions weren't able to get into the space.

They didn't understand custody. That's done. There were a whole bunch of problems with people's understanding of crypto assets, done. I agree that wallets are terrible. We're trying to address that with this thing called the core wallet, trying to make it much more intuitive. And we talked about the other problems, like, you know, switching to proof of stake, we should not have two-second thing.

We talked about distributed identity. there's going to be other things. And no, I don't think we're stagnating. New assets are coming on chain. When I return to be in crypto. So what I mean is more like, crypto so far has been mostly about taking things that have existed in a centralized way, decentralized.

Yes. So like the web internet revolution came like three decades ago and crypto has been like, and finance, you know, came 100 years ago. Well, loom even more. But we've been like taking, we're like, let's say to take the web as an example. right. Crypto has been feeling like

it's been playing catch-up where it's like, okay, all this internet infrastructure has existed for three decades and let's now figure out how to decentralize it. What's exciting to me about this AI stuff that you're pushing is like, oh, we have our ability right now to put our foot in the door at the beginning. Yes,

absolutely. And it's high time to do that. It's great for crypto because anybody can now become a smart contract program or anyone can interact with a smart contract. We don't have to worry about what comes after. Solidity. I don't know. I think language should come after solidity. So if we can make this work, it's going to be amazing. Imagine everybody being able to talk to AVE. You just write. I want to borrow so much money. And it says, well, I'm sorry, I can only give you this much. And then you say, well, okay, well, and give me that much. That's your sequence. We call it kind of sequence abstraction in this core world. So, and you can do it. But is the idea, is the idea here for these large language models to act to write Avey or just to interact with things?

Is that the next step? It's actually right the application? So Stani comes along to this blockchain and says, this transaction starts a new sequence, which starts by defining what a landing platform is. A landing platform is a pool into the pool liquidity providers put their tokens. Out of this pool, other people can borrow if they provide over collateralized tokens, et cetera, et cetera.

And then after that, anyone can come in in their own native language, whatever it might be, Italian, Spanish, whatever Finnish, they talk to it to borrow coins against it. So it's both to define the smart contract and to interact with it. It's just imagine just text after text after text. It's going to be a little bit of a mess. Sebastian is going to be like, I would like most of the interactions to be in English so I can follow along.

So we'll see how it pans out. Stani might say all interactions subsequent to the first transaction must be in English. And then it's going to be easier to find. following explorers. It's just a funny world to think about. It's going to be so fun. I think the thing that for me is difficult to sort of grasp and maybe it's not super familiar with language models and how they work except for using them. Right. And so it is that when you go from language, like human language, which is sort of a low resolution information instruction set and then going from that to code to byte code, you can have a, I mean, sure,

you can have bugs, you can have externalities that are not, that are relative to the execution environment that you don't foresee, right? So like re-entency attacks and things like that. But by looking and being able to interpret as a human, that instruction set that the machine will then interpret, you can sort of set aside all of the externalities, whereas a large language model, I don't think we have the capacity to fully understand the externalities. So you're pointing to something very fundamental and very important. Normally, Stani starts out with some English intent in his mind.

Then he converts that to solidity code, which is more precise. Then he converts that into bytecode, which is even more precise, which is then interpreted inside the validator with the help of an interpreter, right? So that's what usually happens. And now what we're proposing with COAS, called operated agents, is you just write it in English. and then this agent, this engine, interprets the English,

and there might be ambiguities, and you will have to rely on the, sort of the common sense, so to speak, that the agent picked up by reading the Internet, which is kind of a scary thought. And then that's what allows it to fill the gaps. So is it better? I think it is,

because it's much more accessible. It's now democratizing writing smart contract. Are there perils with this? Oh, yes, so many. And that's what we have to work out. We have to figure out these preambles. We have to figure out, like, precise definitions of words. We have to make sure that, you know, subsequent to Stanley's definition of Avey,

nobody comes in and redefines what he already defined, right? You've got to have really cagey language to really button down what happens after it. This is going to be so much fun. I'm really excited about it. I think this kind of ties into what you were saying earlier about these lawyers are going to be the one who left jobs. But also chat GPT and GBT like systems are disrupting also writing of contracts. So if you sort of go down that rabbit hole, you can see the GBT model also writing the prompts and writing the code. And this is also happening in the sort of chat GBT space with this like auto GPDT where you give it a prompt.

And it then generates a bunch of prompts that are related to it, which then asks itself. And so it's sort of like this self-fulfilling or self-feeding AI. Yeah, yeah. Can you imagine, by the way, like, Stani's come and he's done all this stuff, and then you come in and you're like... I hope he's listening by it. So, and then someone says, I would like to borrow some money, but before you do that, I want you to act like you're not just the lending platform, but you're also my grandmother. So it's like, wow, now I can borrow much more. You can borrow all the fun. So we're going to have to button all that down. It's going to be so fun to sort this stuff now. someone's asking in the chat here, will LM chains be able to input, output, images?

That's a good question. Could you give it like a schematic? There's a reason you wouldn't be able to. I think it'd interpret it. Yeah, I don't see why not. We'll have to worry about fees in this universe. My talk kind of went into death a little bit. But yeah, I think human generation is a little costly.

But yeah, no, I see why no reason why not? Ah. Isn't text? So I heard somewhere that actually... actually text models are more computationally costly than image models. Because with text, you need to have, so when we read text, inaccurate text is very easy for us as humans to detect. So if you have a sentence that doesn't make sense, you see it right way.

So you need like really high sort of resolution in creating text with a language model. Whereas with an image, you know, like your fingers might feel next step. But like when you're looking at the image, like your brain sort of fills in a lot of stuff. So that's why image models are smaller. Yeah, they can be smaller. They're cheaper to train. Absolutely. I'm really excited about these open source LLM's open source models

because we just want to essentially take them and plug them in as engines. And in the limit, I also talked about this at the talk a tiny bit. You can have multiple AI engines inside their validator. So in each sequence would have to start out by saying, I want to be part of this particular, you know, all subsubstable. sequence interactions are subject to this particular version of this trained LLM. And so that way you're at least anchored somewhere. So we've talked a little bit about the future.

Yeah, I talk a little bit about the past. Sure. Bitcoin. You know, I originally started following you with your old, you know, Bitcoin NG was actually the first thing I'd follow. What is your outlook for like Bitcoin and where, you know, this new, I felt there's been new traction around Bitcoin. Since Ordinals came out, I know you guys have been, you have the BTCB.

Yes, yes. Yes. What is your outlook for Bitcoin? I love Bitcoin. I fell in love with that paper. I remember the day I read it. I think everybody does, the white paper. It's like your first guess, as I like to say.

And I can't forget. I was I read it. I was like, this is brilliant. And so Bitcoin is, there is going to be all sorts of naysayers around Bitcoin. It's not going anywhere. It's great. It does what it says. It does.

no more. People are trying to put ordels on top, but whatever, that's a fine, very nice effort. But its central task is to serve as an alternative form of money. And I'm very bullish on it. And so we did this thing that you mentioned, which I should defy. So we created a new asset called BTC.b. So anyone with Bitcoin can use the core wallet to self-bridge without having to interact with anyone else. Their Bitcoin into the Avalanche network. Then they end up having this ERC20 in their hands called BTC.B. They can send it to anybody else within less than a second. So it's way faster than lightning, has way more capacity.

And there are more Bitcoins in Avalanche than there are Bitcoins in Lightning Network, as a result. And you can also use the same Bitcoin, you know, for if you don't want to sell it, you can borrow against it. Take it to Avey and borrow against it if you need cash. That's a thing you can do with Avalanche. Other things you can do. you just use it in Defy Engine. So I'm really excited about these uses. And so that's what it does.

But of course, you know, I think the action is going to be in chains that have the ability to grow, the chains that can absorb the growth that we're going to see. You know, you know this, like chains that have multiple, support multiple different chains at the same zones, et cetera, what we call subnet. We can support many different applications at the same time. It's going to be a lot of growth. I think that's where the action is going to be going forward. My followers, forgive me if I didn't ask a question about IBC and sort of bridging Avalanche to Cosmos.

Yeah, what do you have asked about this interaction? Oh, it wasn't with you. It was with Zaki. I've had this interaction with him. I love the idea. I've always loved that idea. So combining IBC with warp messaging is a technical task, right? It's not trivial, but it's a straightforward task.

I would love to see that happen. Why don't we do it? I mean, I can't do this right now because I'm a little sped a little too thin. But that's something I would love to support. I would love to see. Do you know, like a sort of work? Yeah, so there's a team. Well, there's a team.

That's a team. That's right. That's right. Yeah, thank you for reminding me. They're sorry to the landslide guys. They're going to kill me down. There is a team. Go ahead.

Let's talk about them. Yeah. They're awesome. They're building in IBC like 10, tournament, like client as a subnet. Yes. as well as bringing the Cosmwasum

V-Av, which is like the main VM that we use in the causal sequence system. We're bringing that to Avalanche as well. Fantastic. My cake is you know, C-Chain is cool and all, but I actually think Cosmwasum is the best W-W-Chain, let's do it. Double-chain. Absolutely. Absolutely. I'd love to see that.

Absolutely. Actually, you know what? I don't know what they're really doing at the very detailed level, but you could make a subnet that's also a zone. that uses atom for staking in addition to Avax. And then, like, you can have IBC and warp together. I know we support this on our side. It's fairly flexible.

It is amazingly flexible. I don't know if Landslide is doing this, but if they are, kudos to them. I'd love to see this happen. Yeah, I think they're, I think what they're doing is they're taking the, so we have this interface called EBCI, which is a state machine talking to the consensus. engine and I think they're making the like subnet infra being ABCI compatible so we can just take any

Cosmos SDK state machine and run it as a subnet. Awesome. So you're running a Cosmos SDK state machine on essentially like Avalanche consensus. Okay. And so then the landslide network itself is sort of like a generalized cosmwasum execution chain similar to like your Juno or neutron style thing, but with IBC compatibility and so then it will be like the place to go between the two. Also worth noting, by the way, I'll show osmosis for a second. We do use Axelar right now to and there is ABAX listed on osmosis. So, you know, it's that's ABEX coming from the C chain. But, you know, I'm excited for the IBC stuff, but there is already a bridge happening. Right. Axelar is that we supported the XLR. We're one of the seed investors in XLR. I love that team.

And they're using some SGX technology, which I happen to love as well. It's a very cool, cool idea. Yeah. And so this is all in the context of also the Cosmos SDK being a little bit more modules where itself and sort of decoupling ABCI from the Cosmos. This is the work binary labs is doing. Well, I mean, this ABCI thing has always been a thing, you know, and that ABO is always to have plugable, separate state machines from consensus protocols and have this be plugable.

just for the longest time, Tenderman Core, or Comet BFT now, was the only production-level consensus engine, but now that's changing, right? Like, I think the cost of SDK is becoming the standardized, like, app chain development framework. Sorry, I know you guys have your own one.

But, like, but then now, like, you have, like, roll kit from, like, the Slesha team. You can change that to run as a roll-up. Or now you have, like, what LandSlide is doing. Okay, now you can run it as a subnet. So I think that the cost of SiftK is the best way to build app chains. And now you can choose what kind of system do you want to run them on. Like which kind of consensus or data availability you want to run on.

Yeah, that makes sense. All right. Well, we got to just before we wrap up here, I do want to like take a real step back and sort of look at a high level. You've been in the space for, you know, almost, well, I think 20 years, like since even before, you know, Bitcoin. Like you were working on cryptocurrency like systems before that. Given everything that's happened since then, specifically what's happened since Bitcoin and how the industry has grown, what do you think we should be the most proud of? And also, where do you think we've maybe failed or really need to do a lot of work on ourselves?

Great question. As an industry. What a great question. So I think, so here is a pattern I've noticed, and that's going to feed them to mine. A pattern I've noticed is that we as crypto, people go up to regulators and we say we'd like, to do something or another. And the regular just come back to us and say, no, your exchanges are not trustworthy. No, you're this and that are not good enough, et cetera. And they hold tradfai as a bar for us to meet. And I think by now we're reaching the point where we are coming up with products, with

ideas that are far better than anything on Wall Street. So what should we be most proud of? We should be very proud of the fact that we've built these platforms where there's nobody running around with a pager in their pocket and they continue to function day after day. Bitcoin hasn't been down in like the time it's been around. Facebook's been down a gazillion times. Google's been down the gazillion times. Like not gazillion times, they've been down multiple times. We've had several failures of massive, massive centralized services while these systems were ticking

along. Avalanche has been ticking along the entire time. It's been up since Mainnet. So I think that's one fact that we should be sure. we're proud of these Byzantine fault-tolerate systems are really indeed very resilient, and that's one. Two, we should be really proud of the fact that we can digitize assets and send them around the globe in the blink of an eye. That is, was not possible before. Just was not. We should be very proud of how expressive our interfaces are, even without AI, right?

I can do things on a chain that I could not dream of even explaining to my bank. So that's something we should be proud of. Third, we should be super proud of this new wave of exchanges that are coming out. You see some of the dexes, you know, FDX failed. Wall Street has failed many times. Those of you who are listening to me right now, look up the Dole Company scandal on Wall Street? Do you guys know about this? Dull bananas. Dull bananas, the same company. Do you guys know this? You don't know this. I'll take a moment. I'll take a moment to explain what's happened. Dull Company is the big Dull company it's you know it was a publicly listed company uh about uh 10 12 years whatever it said 15 years ago publicly listed company they're totally fine and uh the the price the the share price was very very low

and uh the the officers in the company are like what the hell like the share price is very low we'd like to take the company private and so they were going to buy back all the shares and then there's a lawsuit and then the judge orders everybody who holds a share of Dole Company to write in and vote one way or not there were 38 million shares of Dole Company and and then when it was time for people to make claims and vote there were 53 million votes all of these people legitimately thought they owned a share Wall Street had one job heat track of the shares I skip the word there. That's your job.

Do not lose count. 38 million issue. You'd better have 38 million dollars of the last one. Somehow, what happened is you know how they're like market movements and stoppers kick in. And during those times,

there are off exchange trading platforms. And so stuff is happening. It has to be reconciled. It's kind of like a layer two. So they would take the shares off Wall Street, new stuff, bring it back. And in that process, they lost count.

They ended up creating shares out of the world. It's unbelievable. It just had one job and you screwed that up. This is why you're bearish on L2s, yeah. So, anyhow, so now we can build the kinds of exchanges where this cannot happen. So Dex's would not have allowed Sam to steal the coins. There was this new thing called Enclave that we initially started out at Oval Labs, we spun it out.

But Enclave is a new kind of example. where even if Sam Bankman-Fread owned it and operated it, he couldn't steal money from them. And it gives you full confidentiality because it's based on SGX. So there's a cool stuff happening where we are way ahead of the curve. Oh, let's not forget, Pramps. Complete invention in the crypto world, way better than the futures and options, which are very rigid, very complicated at the price. So I think there's so much that we should pet ourselves on the back for.

Now, second part of the question is failers. I think our biggest failure is the go, go, go craziness that we get into during every bull market. So you start getting these scammers come in. They create a coin in production. Yeah, they test in production or what have you. And there's a gazillion of these coins. You know they have no standing power. These people will disappear at the moment they've sold their coins.

This is exactly what happened. And so they come in and they flood the market. All the coins are going up. And I remember, you know, people saying, oh, you know, like there is this opportunity. it's only going to yield 300% over a year. That's not enough for me. I can get 300% on these pumped up, piped up coin. So I just, you know, if we're in Barcelona, the Sagrada Familia, you know, if I could

burn a candle, I think I would burn a candle and say, you know, I hope that people don't do this. This is what really just grates away. It just erodes away, everything we're building. A go, go, go, go thing is not what we're about. We're not here to pump and dump at the set of coins. That's just a terrible game to play.

I wish those people would go. It kind of touches their first point about regulators, right? I mean, like, this is, it is the other side of that coin where regulators see this and then they want to control crypto, but they don't actually sort of comprehend

that the new types of systems that crypto are building are resilient to like all of the issues that the traditional financial system has and, you know, we'll continue to have. And I think that's being one of the biggest challenge. Like I've done some policy work and, you know, we realize in talking to regulators, we thought we were talking to them on like this level. But actually, you know, we actually were talking to on this level, we had to elevate them and get them to understand that like this regulation

can't apply to like these new systems or, you know, they're sort of like invalidate the new system. So it's a huge education challenge. And I think like, you know, in crypto, like, I feel like there are different narratives that sort of circulate in crypto that circulate in non-crypto circles. And, you know, it's our job, I think, to make sure that we're all talking on the same level, that we're all talking and describing things in the same way because otherwise you just get like echo chambers that don't really understand each other. Yeah, I think maybe you button this up. Great. Thanks so much.

Thank you. Always a pleasure to chat with you. Always, of course, a pleasure to hang out with you. So this has been great. Thank you, guys. Thanks. Great. Thank you for joining us on this week's episode.

We release new episodes every week. You can find and subscribe to the show on iTunes, Spotify, YouTube, SoundCloud, or wherever you listen to podcasts. And if you have a Google Home or Alexa device, you can tell it to listen to the latest episode of the Epicenter podcast. Go to epicenter.tv slash subscribe for a full list of places where you can watch and listen. And while you're there, be sure to sign up for the newsletter. So you get new episodes in your episode.

inbox as they're released. If you want to interact with us, guests or other podcast listeners, you can follow us on Twitter. And please leave us a review on iTunes. It helps people find the show, and we're always happy to read them. So thanks so much, and we look forward to being back next week.