Loading summary
Transcript85 lines
- [00:00]
Balaji Srinivasan
All right, Sean, welcome to Network State podcast. Glad to have you here.
- [00:04]
Sean Neville
Great to be here, thanks.
- [00:06]
Balaji Srinivasan
So we worked together several years ago on what became usdc and I was the lead on the Coinbase side and we set up the Center Consortium and you and Jeremy were the co founders of Circle. And I think the Circle Coinbase center partnership has been pretty, pretty impressive for the world over the last seven, eight years. Do you want to talk about that? Give. Just, you know, introduce people because, you know, you're a co founder of Circle, you've done a bunch of other things you could give the. Give the quick spiel, Sean. On Sean.
- [00:34]
Sean Neville
Yeah, yeah. So Jeremy and I started circle in 2013. So it's been, you know, 12 years, which is like 120 years in crypto or something. Yeah, exactly, yeah. And so, you know, we were, we were just obsessed with this idea about democratizing and sort of decentralizing global finance and sort of had this vision for what that might look like. And the vision really hasn't wavered. The way that we've tried to execute on the vision obviously has had a lot of different sort of crisscrossing paths and probably should have been obvious in 2013 that we needed something like a stablecoin, but it was actually 2017 when we figured out that's what we needed in order to build everything else that we were interested in. And yeah, I was going to say, obviously, you know, that part of the story is as you came into the picture and helped make it happen as well.
- [01:25]
Balaji Srinivasan
Yeah. And I think basically what was funny about that was that was, I think that was true teamwork. And what was interesting about stablecoins at the time, as you may remember, is a lot of people poo pooed them and said, wow, your great innovation is putting a dollar on chain. That's a great innovation. Right. And it was kind of like people not understanding how big a deal it was to put like for example, news on online. Right. Or, you know, they, they just didn't understand that for example, once it's on chain, you've got programmability and you've got, you know, anybody can send to anybody and you can use it in smart contracts and you can slice and dice it down to smaller amounts and you can have devices use machine payments as we'll get to with your new startup with the Katina. And people didn't understand that stuff at the time, but I think we saw that. And what were you thinking about? You said the vision had changed with stables over time, but what was the number One thing that you guys were thinking about at the time you started work on usdc?
- [02:28]
Sean Neville
Yeah, I mean it was pretty simple honestly. In order to enable global payments that didn't pause at borders and that were almost free, we needed some representation that was stable and capable of moving over Internet rails. So the idea was multiple chains, not to be religious about any one particular chain or piece of infrastructure, but to rely on blockchains to deliver value. And that in itself was really a building block so that we can enable these other use cases. You know, I'd say today stablecoins, you know, largely the use case has been in crypto capital markets. We haven't yet seen that unlock of payments, treasury management effects, you know, these other things. But we're right on the cusp of it now. And you know, we certainly needed that building block technically. But also it couldn't just be a circle coin or a Coinbase coin. You know.
- [03:15]
Balaji Srinivasan
That's right. That's why usdc, if you recall, I think we named it so that the C was circle and Coinbase and Center because the center consortium and the consortium meant it was decentralized because we were partners on it. And it also meant that others could in theory join.
- [03:31]
Sean Neville
That's right. And that was the idea. We expected there to be multiple, you know, potentially multiple issuers, all of whom participated in a consortium, even though they may be competing at other layers of the stack. Um, you know, it's the, the idea is just interoperable standards. Right? So like back in the day, Microsoft had a slightly different version of HTTP than, than Netscape, ultimately.
- [03:53]
Balaji Srinivasan
That's interesting.
- [03:54]
Sean Neville
Well, you know, there was this, there used to be, it was, it was sort of like, you know, build your service that ran in ie.
- [04:02]
Balaji Srinivasan
Do you mean like blink tags versus Marquee, like HTTP, that kind of thing?
- [04:05]
Sean Neville
Exactly. Yeah, yeah, yeah, yeah, yeah, yeah, yeah. But ultimately everybody agreed. Let's just, let's just agree on a standard and let's not argue over different implementations of say ssl because we can compet E commerce layer so long as we have this foundation of interoperable standards that no one vendor really owns. And when it comes to, you know, delivering dollars or other currencies. But stablecoins today largely, people largely want dollars today is what they mean when they say stables. It can't be circle dollars or stripe or coinbase dollars. Yeah, exactly, Coinbase dollars. Whatever it is, it's just dollars to people who are using it. And so, and so, you know, we sort of saw a need for if dollars were going to Move over Internet Rails. Then it needed to move as an interoperable standard. And so that was the idea. Put together a consortium, not sort of libra style of 30 people who can't agree on anything, but a small number of very like minded, motivated participants who could agree on these fundamentals and get it going.
- [05:07]
Balaji Srinivasan
Yeah, you know I like the Libra guys and I remember saying to them at the time and I think David Marcus and Morgan Beller and so on would agree with me now I guess it was totally fine, they did their thing and again them props for seeing it through even though you know it didn't work. It not everything works. But I think the issue with Libra at the time was a, they were using a basket of currencies to back their thing. And the problem with that is you had volatility without upside. Right. You didn't know if it was like one Libra is what it's 87 cents today and it's 15 tomorrow. Right. And so it wasn't mapped to anything that people had psychologically, right? No, no fixed price. Yet at the other hand they had volatility without upside because it's never going to be 10x. Right. So it's like worse of both roles in that sense. And then there are other kinds of things that were like that as well where it was blockchain so traditional banks thought it was risky but it didn't have any upside. So crypto people didn't want to back it. And actually this was something which I think we did right with usdc. So I'll explain on the coinbase end how we were thinking about somewhat similar to how you were thinking about it. So my view was because at that time, if you recall, we were in the ZERP era, right? And so we didn't know, I mean only with a crystal ball would you know that they would jack rates to the moon in 2022. Right. That was like four, four years out at that point. And we had been in ZURP for like almost 10 years, right. Since the 2008 financial crisis. Right. And so USDC, it was not obvious that it was going to make any money anytime soon. And the way I conceptualize on the coinbase side is we thought about it similar to how you thought about it, but we thought about it as Google login. Like Google did not make money directly from Google login, but it sort of projected, you know, like projected power, projected login over the whole Internet and then people would come back to Google except a decentralized version of Google login. Where stable coins would be used everywhere and then people would come back and they'd trade on our venues, they do things. They'd basically be part of the crypto economy. So just like Google login sort of grows the Internet, we thought USDC will grow the crypto economy. That was one kind of process. The second thesis is we, we put in capital behind it because we didn't think that. So had it been an ico, for example, there would have been pressure for a short term return. So we capitalized it, you know, together. And I think that proved correct because had people gotten or wanted interest back from it. Now today it throws off whatever billion dollars a year. It's like at 5% rates, 4% rates, 70 billion, it's like 2, 3 billion dollars a year. Not bad for seven years. Right. But at the time it literally returned like nothing for like the first four years or something like that because interest rates were zero. So something like that is, you know, something where you can't necessarily make money on it because rates are not, you know, we're not within our control or we didn't know. So we had to essentially assume that it was just a utility thing for the entire ecosystem until the returns came there. Let me. Does that comport with how you, you guys thought about, know your thoughts on that?
- [08:06]
Sean Neville
Yeah, you know, the first version of it in 2017, actually predating Coinbase, we did have a, you know, it's probably still floating around out there. The first version of the white paper contemplated a token being released on the network that was called the Scent. And, and so there was sort of this, this alternative path that we quickly realized this is not the right way to deliver this. But there was a little bit of a different revenue model tied to it by virtue of, you know, having the token. And this was, yeah, this was in the ico, you know, sort of boom and bust. And you know, we explored a lot of things to try to deliver it in that fashion before we did V2.
- [08:46]
Balaji Srinivasan
Yes, our collaboration. And then we, that's right, we, we, we had patient capital on the balance sheet. So we didn't need to, we didn't have ico. Like nothing against coins, but those investors are not patient. As you know, ICO investors are not patient. So, you know, so by having patient capital on the balance sheet, we could go for years and years without, you know, worrying about any return in the short term. And I think that that helped a lot. Okay, so now I'll tell you before, before.
- [09:17]
Sean Neville
So, you know, we had sort of decided, well, coinbase is the right partner. We were also talking to others, I mean, as you know, you know, and assisted with like, by people like Jim Migdal at Coinbase. We're trying to bring people into the central consortium. And there were very serious conversations before Coinbase that just, it didn't work out with very large, you know, players when ultimately you, you leaned in and others at, at Coinbase, I always wondered internally, did Brian have to be convinced to, to, to go this path?
- [09:49]
Balaji Srinivasan
So, great question. So, so now, now the story can be told, right? So it actually. So again, of course, you know, everything that happened on the Circle side, I can describe what happened on, on the Coinbase side. So there was essentially a big conversation internally, which was build versus buy versus Partner. Right. And I wanted to, so buy would have been like, buy an expensive stable coin or something like that. And I also felt that wasn't really in tune with crypto, because crypto is about decentralization and so on and so forth. And so I wanted to. First of all, partnering would have been. You guys had my view on what you guys had. You had an audited smart contract at a time when that was actually a relatively very scarce thing, right? Because a trail of bits and what have you had, you know, they were, they're still, I think they're still good, but they were backlogged. And, you know, there weren't that many people who knew how to do audited smart contracts at that time. So I knew that that would take us months to get to the level that I felt that we could put hundreds of millions or billions of dollars into it. Right. Like an audit smart contract is no, no small thing. Right, there's that. And also, you know, the fact that you guys were reputable, you'd been around for a long time. Jeremy had had an exit before with his video company. Gosh, I'm forgetting the name of it. But it was. Exactly. That's right, look, Streaming Stream Video company. And actually we had met you, I don't know if you remember this, when I was partnered, Anderson, all the way at the very beginning in 2013 when you guys were starting Circle, you know, so, so we, we, so I, I, we had game film on you. Right? And, and obviously I think there was mutual respect. And then the, so my view was a partner, B, use our balance sheet. And essentially the question was partner versus Build. Right. And I felt that partner was the right approach because crypto's in like, I'm a, I think I, I like to be. Or we'd like to be win Win people and, and, and there's more that can be done together. And I felt that that would be an organic partnership since you guys were also crypto first and you've been in the space for a long time and everybody who's in the class of 2013, we all have our grays. Right? But that was like the first time that a bunch of know, Eric Voorhees of Shapeshift really got his start back then. Right. I think blockchain info really got going then you know, Blockstream got funded around that time. So I think quite, quite a few companies as you may recall, like Coindesk and so on really got going. So I think some of the really foundational companies in the space got going at that time and we, you had been in the space for the right reasons. You weren't like a Johnny Come lately and I mean it's all obvious stuff but these are just explaining what, what our thinking was. So I felt you'd be a good long term partner for this. And so that's, that was part of the reasoning on, on our side and then what I, what I did, which you, you guys don't care about. But we, we had just really three or four people on our stage. Jim McDowell who you met, Jacob Horn who's the product manager and then Maxim and Miha who are the two lead engineers. And then there are several other people, you know, on design. And so for those are I think the four principal people. Maybe there's somebody I'm forgetting, but I think there's the four principals. And the biggest, one of the hardest things was just clearing everything off of Maxim and Miha's schedule so they could just code 247 and it block all the HR people and so and so forth. So no, I just fill out all their forms and we shipped it really fast. 10-23-2018. Go ahead.
- [13:08]
Sean Neville
Yeah, yeah, I think that's right. And Jacob by the way also I always felt he was a great champion for us internally at Coinbase and you know, really, really enjoyed working with him. We had Joao Regan, it was a PM on our side, but our, he is really small too. It was a handful of people. I mean the tech obviously it needed to be secure, it needed to be a strong. Obviously we launched it on Ethereum. It's now on like 20 plus on everything. It's on everything. But it has to be secure. And so the code mattered a lot. But there's not a whole lot of tech there. I think the partnerships and the distribution obviously trying to clear the Liquidity moat which is still, you know, there are only two dollar stables that have any meaningful liquidity right now because that liquidity mode is really difficult to clear. Like all of these things were at least as important as the pure engineering.
- [13:59]
Balaji Srinivasan
That's right. And I think the reason on the engineering side it wasn't so much it was hard because it wasn't a lot of code, but it had to be extremely correct because you know, smart, you know, reentrancy bugs, there's all kinds of these things that can get you. There are two other things that I remember at the time that were really. First was that a lot of other people had been doing stable coins backed by what I call plutonium. Right. A very unstable thing over here. And you didn't know what the value was because it jumped up and down all the time. And nothing against tethered. They've proven themselves, they've executed over the last X number of years. Right. But at the time it wasn't known that you know, they're, that they're. They had enough paper to back with their outstanding assets were and so on and so forth. Right. And then there are other people, not even tether who are doing algorithmic stablecoins, some of which as you know, you know, blew up later. And so one big thing for us was using the banking relationships to get $1 on chain backed by $1 off chain and do it in a, in a dumb. But like, you know, simple is hard and good. Right? Like, you know, so the simple thing of just one currency off chain, one currency on chain that was one big piece. The second big piece, if you may recall is, you know, and this is a big debate with those banks, right. Was would people have to KYC to send to somebody or would it be blacklist rather than whitelist? Right. Meaning would we maintain. So we had to have some compliance type stuff and so on and so forth. Somebody misused it. So we eventually agreed that freeze and you know, in, in, in extreme circumstances freeze or reverse transactions in, in the contract was better than having every single address have to pre verify in KYC beforehand, which would defeat the whole point of it. Right. So once we, I think that was another key thing is we got enough banking partners to agree A okay, with a warrant you can freeze or seize or, or, or reverse transactions just like you can with wires. And B you can also guard the egress and ingress at the exchanges when USDC is being swapped for USD. Right, right. And so given that then you otherwise have fun on chain and do what you need to do. Right. And I think that actually now opened up the door for, among other things, something you and I have both thought a lot about kind of your new company, which is Katina. Right. Because by doing that, since you know, if you had to KYC every address, you couldn't have a program generate 100 new addresses because it all have to be KYC by like the human and some photograph or something like that. Too much friction on be impossible. Right. So that brings us to the unlock of machine to machine payments. So why don't you talk about your new thing, Katina?
- [16:50]
Sean Neville
Yeah. So I also say something on the way that the sort of funds freeze happens. I was still never, honestly I always felt like we could have, we could have innovated there a bit more past just address management. And you and I talked at the time about things that were more sophisticated, probably would not have been executable then, but this like governance contracts that. Governance contracts? Yeah, like risk and reputation on chain to, you know, sort of handle these sorts of things. And it turned out that you were right. The simple approach was, was the one that was going to work, you know, but I still always felt like, you know, there's something there and it's the, you know, one of the things that always nagged me about the Internet is there's no real like, like really, you know, distributed identity layer where people can sort of manage their own credentials effectively and not have to lean on a Google or whatever. And I was sort of like the remnants of that. I wanted to somehow figure out maybe how we could incorporate it. But at the end of the day it was all about not even getting the regulators comfortable, but our de facto regulators, which are our bank partners. And this, the simple approach was the one that was going to win there for both settlement and reserve bank partners. Yeah. Yes.
- [18:03]
Balaji Srinivasan
And I think that basically, you know, like you. So I think today it's possible to innovate on that. Right. But you know, my, my view on a lot of things like this is sort of like minimum necessary innovation. Right. So no coin, we just did off the balance sheet, no plutonium, we just did one to one backing, you know, USD for usdc. No COIN governance. We just had root access to the contract. We had the minimum size of the consortium of the game, which was two. And we just minimized all complexity and it was still not a completely trivial thing to scale and build this thing over the last few years. But it was awesome. And really, it's funny, you do 20 things and it's interesting as to what things actually turn out to be really big. It's kind of like for, you know, they're like 50 investments I do did in a year and then one of them is like Ethereum or one of them is Salon or something like that. And then it's like, oh, okay, well that went, you know, that goes really big. Right. And it's interesting because a year later or two years later USDC was useful. But it is, it is after I think Zerp ended that it really went boom like this and became as material as it is now. Right. And it's funny. So I'll tell you another discussion that not public, but I think, you know, there was a discussion internally as to how much to out, you know, like resources to allocate towards usdc. And unusually, see most of the time, you know, good advice at a company is like ignore macro. But in this case you couldn't ignore macro because 100 of the decision was what is the Fed going to do? Right? Because it's like is it going to increase rates or is it going to keep them down or is it going to go wiggle like this or what? What is it going to do? And like it felt actually it was much more Wall street than I'm used to in tech because it's completely 100% decision was your mental model of the Fed. I don't know if you have any thoughts on that, then we can go to Katina over there.
- [20:05]
Sean Neville
Yeah, yeah, yeah, yeah, yeah. No, that's true. I mean, I think that my recollection is that even after we had, we had decided on most of the structural pieces in terms of governance and tech and we had, you know, lined up the banking partners. And this was also, speaking of bank partners, this was also just barely after the era when there was only really one bank in the world that would bank coinbase or circle.
- [20:30]
Balaji Srinivasan
Yeah.
- [20:30]
Sean Neville
So the banking partner piece which you mentioned was, was a huge negotiation and discussion because things are different now. You know, we have, you know, Bony Mellon and BlackRock and you know, you know, but it was not like that then a short time ago. So. Yeah, but my recollection is the most, I wouldn't say difficult but time consuming, hardest pieces to put together were the economics, it was the macro situation sort of analysis, but also micro. Is this. What should the economics be? If we have a consortium not of 100 people but of a small number starting with two, what does that look like? And there's a lot of discussion that's been revised since our time focused on that but, but yeah, absolutely. I think that that was the big puzzle. And even then I remember we had people talking about, you know, maybe can you do JPY stablecoin or like negative interest rates or this didn't, you know.
- [21:27]
Balaji Srinivasan
All that is now happening. All that is now.
- [21:29]
Sean Neville
Absolutely. Yeah.
- [21:30]
Balaji Srinivasan
Right.
- [21:31]
Sean Neville
Yeah. So, so that's right. So, so you asked about Kapena and I keep going back to USDC because it's fun to chat with you about it. But sure. You know, one of, one of the nice things about, about machine native money is that it's, it's very reliable for machines to use it. So when we have these orchestrated workflows, no one can agree on what an agent is, but sort of workflows with LLM intelligence involved in making decisions, then as those workflows become economic actors, they need to be able to transact, they need to be able to hold money, they need to be able to send money, they need to be able to make treasury effects decisions potentially all within guardrails and humans in the loop where necessary. But simply tacking AI intelligence onto the edges of say a credit card flow is just aside from the economics, it is not reliable. It's really difficult to make those even just the simple tool calls very reliable. And so when you take a five to seven to nine party system and you add a couple more parties on the edges of it, then it's different than taking a stablecoin point to point transaction. Agent tech workflows are very good at signing cryptographic messages messages and then all of a sudden you sort of unlock the capability of transacting in fractional amounts in streaming payments in real time. All these things we've been talking about for 20 years that have really never been economically possible or technically possible until we have the combination of machines becoming economic actors and stablecoins. I don't know if we keep calling them stablecoins, but it's screaming money fiat.
- [23:12]
Balaji Srinivasan
That's right. Well this is funny because 10 years ago I also took a crack at the machine payable web and machine micropayments and 402. And the issue was that Bitcoin at the time, see Gavin Anderson, if you recall, had published a roadmap for scaling to huge numbers of transactions of big blocks and so on. And then the bitcoin civil war happened and became small blocks and fine, but that just basically meant that you couldn't scale on Bitcoin in the way that, you know, lots of transactions so everything had to move to other chains. And so, and frankly, you know it's funny now today in retrospect, now that we have Solana and Base and we have USDC and we had to have Ethereum and we had to have Bitcoin and we also had to like win this gigantic political battle right after all of that now finally lots of the pent up innovation and also we had to be able to get on phones right, because we had to win this political battle because this is something, it's a really interesting thing which you know, and I know so many people for so many years are like where are the crypto apps? Why aren't they on phones? And the reason is that Apple especially and to a lesser extent, but real extent Android would nerf all kinds of apps that use crypto on phones because there was something like Apple payment, you know, the IAP thing in app purchases thing, all this kind of stuff where they've now actually lost some cases on that, you know, and they started being forced to do that. And to be clear, look, Apple does a lot of stuff well, but they're sort of genetically anti crypto. And so there's a chicken and egg thing where you just couldn't get. And that's why crypto is actually developed as much more of a web phenomenon than a mobile phenomenon. We have to be choke pointed in a certain way where we're only web rather than mobile and we have to be choke pointed through the banks in a certain way. And like you couldn't do equity issuance, you can only do meme coins. So I think Bankless also had a very similar observation like there's all this pent up spring of innovation go boing like this over the next 10 years that, that, that machine payable machine payments.
- [25:12]
Sean Neville
Are a part of.
- [25:13]
Balaji Srinivasan
Go ahead.
- [25:14]
Sean Neville
Yeah, yeah, absolutely. I mean my thesis is that in the future the only actors that we will trust with our money and our assets and the only actors that will be capable of generating competitive returns will be agentic.
- [25:30]
Balaji Srinivasan
Interesting.
- [25:31]
Sean Neville
And I mean.
- [25:34]
Balaji Srinivasan
Or do you mean delegates of yourself?
- [25:36]
Sean Neville
So this is, that's a good question. So right now there's a little bit of a, of a battle I would say between you know, those who believe that agents will always be, should be mapped to tasks that always get mapped to some user could be, you know, business or consumer or whatever. And those, those that believe that may be true, but we'll also see semi autonomous or autonomous agents that actually do have their own identity and act on their own, not just a planning loop, but can truly act on their own behalf. And so there's debates on, on both sides of those things. But I, I think you know, either, either way, you know, in terms of being able to make good decisions rapidly enough with, with money, we will just not want to trust anything other than. I mean this is hard to imagine now because people's anecdotal experiences are that, you know, Claude or ChatGPT can't give them a, you know, recipe for a brownie correctly because of the sort of hallucinations. But it, but it is, I still have strong conviction that it is the truth that we will see AI actors that will be the most effective economic participants the world has ever seen and then we will want to use those but we will have no need to ever execute a transaction ourselves. We will always be doing it through some form of personal or other AI.
- [26:51]
Balaji Srinivasan
Interesting. You don't think you'll ever need to execute a transaction yourself. That's interesting. That's a strong form version.
- [27:00]
Sean Neville
So where do you now when that happens? When that happens is a question I always find the when the hardest question for me to answer in this. So I have conviction around the what does it take three years to touch certain domains before others or does it take two months? Does it take 20 years? It's really difficult to say.
- [27:18]
Balaji Srinivasan
So 12 years ago actually me and actually the Winklevosses and novel there's some panel and I remember thinking about machine payments then and my kind of example of something Bitcoin could do that other currencies like Fiat couldn't do is if you had two self driving cars on the highway, one of them could pay the other to pass for example, right. And now that's just kind of a made up example. But the general concept is machines could negotiate prioritization amongst themselves or what have you. Right. In practice by the way, self driving cars might just drive fast enough that they could just, you know, boom, snap, snap onto each other and, and, and go. Right. But there's probably something to that of machines being able to negotiate back and forth. You know, for example you could have a fully, you know right now Waymo and, and Tesla Robotaxis just to drop off and pick up. Finally we have truly fully self driving end to end cars. But you could have something load food into it and then unload the food on their side. Like sidewalk robots exist, right? And drive throughs exist and you could have something where like the car pulls up, it pays the robot, it gets the food. You know what I mean? Right. So that's an example of machine to machine. I don't know if now the thing about that is that's actually in the physical real world, which is always hard because there's other kinds of issues. WI fi dropout, the Bluetooth drop. So I wanted to know where you thought the first two or three applications of machine payments would be for Catena.
- [28:47]
Sean Neville
Yeah. So I would say what we're seeing today, what we're not seeing is true agent to agent or machine to machine payments. We're seeing AI workflows pay for access to resources, content, data, paying for access to APIs. There may be another agent on the other side of an API, but the interface is not truly agent to agent yet. And I think there are a lot of reasons for that that have nothing to do with the payment side. There's no DNS for agents that we haven't really figured out the right way to handle discovery. If your agent interfaces, you could just.
- [29:21]
Balaji Srinivasan
Do it all with text fields, I guess. But yeah, not really because agents don't have persistent domains either. They can't. Yeah, I mean, they kind of do, but not really.
- [29:29]
Sean Neville
There are proposals for like registry concepts, who. Google has, you know, proposed something, a registry concept attached to their eight framework. Maybe, maybe there isn't really like a standard for it.
- [29:39]
Balaji Srinivasan
Maybe you could use ENS or sns, the Solana name system for doing that because that's a little more flexible than DNS. Yeah, you can buy.
- [29:46]
Sean Neville
I think we may see a little bit of fragmented approach, whether it'll be, you know, multiple things that will need to be supported. But in the meantime, it's sort of like, you know, if you and I go build a website, we get a domain and we just deploy it or, you know, there are lots of ways for it to be discovered in the world and there aren't really, there's not really that equivalent for agents working outside of their ecosystems. And even when people build agents, it's usually in a particular framework or in a particular ecosystem. It's. And it talks to other agents in that same ecosystem, but it doesn't really span the ecosystem. So if you build an agent that you deploy, say you build in, say, you know, a framework like Lane Graph, how does it make itself discoverable outside of like a chat interface to an agent deployed in like Salesforce or something?
- [30:29]
Balaji Srinivasan
So, so, you know, there are smart contracts that are something, something.eth. right. And so that, I mean, basically the smart contract and the agent are the same thing potentially, and you just have an ENS name or an SNS name. And that is, that is the registry. But, but I like Your framing of it, which is smart contract registries and agent registries should be the same thing. And that there should be like, I think that itself is a business in its own right to just register and discover agents, you know.
- [30:59]
Sean Neville
Yeah, I think that there's probably a, there's some form of, you know, verisign for agents company that wants to be.
- [31:06]
Balaji Srinivasan
Built that is verification on them too. Yeah. Because they could just take your funds and something. Yeah, exactly.
- [31:11]
Sean Neville
Because the real, and this is part of the thing, part of the work that we have done at Container so far is just this foundational layer of trying to address trust and policy. The issue that people have with agentic workflows is not yet things related to price sensitivity or even largely data privacy issues. It's just reliability. How can you trust these things? Who's vouching for them? Or when they exceed their guardrails, who's liable? All these sorts of guardrail policy reliability, really trust issues. And so this is why.
- [31:43]
Balaji Srinivasan
Yeah, exactly. That's the thing, the AI, like, you know, I recognize that it's in like in some sense a trough of the Gartner hype cycle. Now it's weird because both there's tons of money going into it, but also there's less energy than there was a year or two ago in the sense of oh my God, it's going to kill everybody and so on and so forth. Right. But my view is at least right now, as of 2025, AI requires a lot of supervision to get anything non trivial done. And the smarter you are, the smarter the AI is. And it doesn't do it end to end, does it middle to middle. Right. Because the prompting and the verifying have to be done by humans. And so it's great at generating reams and reams of text, but often a lot of that is filler. And I can always tell when someone's used AI to do something and even when you're using like I think the biggest thing for me whenever we use AI internally is there's actually only relatively few kinds of tasks which can truly tolerate non deterministic error prone output. Images and video can, but back end code can't. Like front end code is more tolerant of those issues because we have our GPUs in our eyes and we can instantly see it and verify that the widget is off. But back end code is much more subtle and it's like much harder to determine that it's off without like really going line by line, you know. Right. And it's possible there's some like visualization mechanism, you know, like a Fourier transform or like an audio spectrogram can turn something that's not visual into something visual. So it's possible we might be able to turn some of these other data structures, like backend code into things that we can just visually debug by eye, you know, like a state machine or something like that. Right. So it might be, might be something along those lines or it's in a domain with smart contracts, for example, as you know, formal verification finally became useful because these were such compact programs and of such high value that all the compute for formal verification actually became valuable. Right. So maybe that's the answer is, you know, we, we have it only generate code that we can do formal verification on or something along those lines. But the reason I ask is where, you know, shopping seems to me to be like the simple kind of thing of get me a good plane ticket that satisfies my requirements like from point A to point B that doesn't cost more than X seems to be like a good place to start because that actually does take a surprising amount of time to go and book tickets. It's like actually kind of a pain still, right? Oh, you have to put in your name, blah, blah, blah, fill in all this stuff on every new, you know, airline site. So that's like one thing, right. And I think the max budget stops it from totally screwing up too much. Right. And maybe you just assign it to get food or a book you like or something like that. I don't know. Maybe you have some thoughts on what it would buy for you first or what you would do with it first.
- [34:41]
Sean Neville
Yeah. So my belief is that right now, I mean things could change by the time this conversation is done. This space is moving so quickly.
- [34:50]
Balaji Srinivasan
Right.
- [34:51]
Sean Neville
But right now I think consumer retail shopping experiences will be one of the hardest nuts to crack in terms of a full shopping flow. So I think I can see it replacing search, which is sort of already happened to content on the web in some way. But the full flow, I think that'll be, I think we'll see B2B or even B2B2B. Sort of agentic payment solutions happen before we see some of the consumer use cases.
- [35:19]
Balaji Srinivasan
But how about EC2 then? Compute auctions, right. So Amazon has auctions of compute because they've got real time pricing as demand goes up and down. Right. And you can get reserved instances versus real time. Maybe somebody's already doing that. But that's an example of a machine bidding on a resource used by A machine for an algorithm where it's a, it's like a truly a machine economy. I don't know, maybe.
- [35:42]
Sean Neville
Yeah. And use cases like you know, supply. Supply chains are another one where a buyer agent and a seller agent could potentially negotiate. So If I have 12 approved vendors for this piece that I need to acquire then this sort of discovery negotiation and also handling things like the fulfillment could be handled by agents on both sides. And in that case the data may be known but in many cases the data types themselves, the schemas may vary and so you have this need to parse not necessarily unstructured, but not always, you know, schema compliant structured data which LLMs can be quite good at. But to make that reliable, you know what we found so far to make that reliable, it's not the AI that is actually executing the task or interpreting the data and beginning to handle the negotiation that requires human subject matter expert alignment. It's actually the evaluating sort of LLMs jury sort of evaluators that are other AIs just evaluating the output of those tasks executors. It's that layer that requires human subject matter expert alignment in order to make the sort of upstream task much more reliable and effective.
- [36:56]
Balaji Srinivasan
So actually just poking on that for a little bit. The only thing about supply chain is I actually ordered a lot of. I still building network schools so actually I'm constantly looking at bill of materials and so on and so forth for various physical things. That is a manual in my view a fairly manual process because you're talking to the vendors you are. I think an agent can help you with getting quotes. Maybe that can be helpful, right. If it's hitting a bunch of telegrams and whatsapps and so that that actually could be quite helpful. Like I have a spreadsheet and I'm just like agent, get me give me quotes on this stuff where it's like only partially listed a little bit like OTC markets, you know, where some stuff isn't listed that's maybe helpful but a lot of those are like large enough and I also don't need it so fast that I wouldn't review the purchase order before I hit submit on it, you know. But that example with like EC2 compute is something where you might literally put that in a subroutine of get me the best price on this which people already do for reserved instances and so on, you know. So like that's kind of machine resources like getting storage, getting computer that feels like where it's like kind of native, you know or I Don't know, maybe. Go ahead.
- [38:12]
Sean Neville
Yeah, I think that, you know, as the sort of business model of the web transforms from, you know, search and advertising dominated into, you know, sort of data curation, this sort of data curation flywheel that is beginning to emerge, which is intelligence. Whether it's agents or LLM foundation, the people providing agentic intelligence need data. That data needs to be curated ultimately by humans. It can be synthetic data, but it ultimately needs to be curated and sort of aligned by humans. And humans need to be paid for that job. So that creates a little bit more of a data content service marketplace that has not existed. And the services that are helping to provide that data to the intelligence providers will pay humans for that job. And so the sort of task of the human begins to be curate these, this data content, these services so that the intelligence ultimately is better for us. We can be more productive and we get paid for doing that task and then we pay for the intelligence in turn. And so whether it's machine resources for things like compute, or whether it's particular content or data or services that can be quantified in like a data schema or something, then that becomes a place where agents are passing money, making payment decisions, but also actually executing, you know, a binary payment task. So very cool. The thing that, where we sort of move, you know, a little bit up the stack, so we're still missing some primitives, by the way, to make that reliable. Some of them are related to this potential certification or identity layer that's necessary for agents. There are many people who are working on different versions of that. But it's kind of notion of agentic identity and handling authentication effectively. Even in the shopping example, if I'm doing, if I'm shopping at Amazon and I'm not just using a chatbot, I'm not on Amazon.com, but I'm just interacting with the Amazon agent in some other way. How can I be sure it's Amazon? There isn't sort of this like, yeah.
- [40:08]
Balaji Srinivasan
The authentication, the verification. Yeah, it's like, who's, whose dog is this?
- [40:14]
Sean Neville
Whose dog is this? And then, and then, and then, you know, what are the new security and risk factors that emerge when AI actors are the ones who are executing these kinds of transactions or an agent on my behalf and an Amazon agent on its behalf. What new risk vectors emerge? This is why ultimately we decided to create a new financial institution from the ground up. Up using AI is because, and we know this from circle, you know this from, from, you know, building infrastructure to manage Risk as well. You know, classic finance and banking risk infrastructure is designed to make sure no bots can ever use it. You need to be a KYC or KYB individual. What we actually need is completely the other way around. Let's assume the only actors who will be using this system will be bots or agents. And, and how do we let the good ones in? How do we keep the bad actors out? How do we give them rules and so that they, they can't exceed? How do they, how do we tell them to escalate to us as needed when they do exceed those policies? How do we have insurance or sort of liability protections for this? There's all these like these problems have not been cracked. And so ultimately we decided, well, I.
- [41:22]
Balaji Srinivasan
Mean self driving cars will be the first place where this stuff is happening now in some ways liability, all kinds of who's at fault, all that stuff.
- [41:28]
Sean Neville
Who's at fault? Exactly. And there is a path obviously to have sort of, you know, we're mostly at level two, I think there may be one level three self driving in the US now. But anyway, we're mostly at level two. There's sort of a path to get, you know, level three self driving. There isn't really a similar path for agents executing financial transactions or you know, aside from commerce doing things like handling treasury management are really large sums of money. And that is where we're headed.
- [41:54]
Balaji Srinivasan
Yes, well, I think you might want to, I mean obviously agents and machine payments in crypto, where an agent is trading on crypto bots, that's like a place where you start with yeah, some, sometimes small, sometimes large amounts of money so that like people have been building some intuitions on this for a while. Right. But, but I think you're right that we could do, we can do a lot more. Okay. I want to, I wanted to. Unless you had something else on this. I wanted to switch gears a little bit and just ask about some different topics, if that's all right.
- [42:22]
Sean Neville
Okay, go for it.
- [42:23]
Balaji Srinivasan
So, okay, so what, what I've been thinking a lot about, what I'm working on is physical world crypto in a sense. Right, because you have, I think after cryptocurrency there's different directions you could take it. But I think crypto community is a big part of it because we go to all these conferences and a conference in my view is actually like a, it's an important subroutine for our space because we're so digital and the conferences are where actually we all kind of come together in the physical world. And so we've got this DARP society here off the coast of Singapore where you've just taken over an island. We've got thousands of people from around the world. It's actually really cool you should come and visit. So I wanted to know what, you know, what do you think about startup societies, network states? Where's Sean on that? Have you given any thought to it?
- [43:11]
Sean Neville
Yeah, I mean, I think we're already in a case where many of the things that even 20 years ago were done in a confined sort of geospace, they're online now. And so we've already, whether we want to formally sort of give a name to it or not, we have informally formed our own communities and our own societies and we're in multiple ones of those with different identities that we present to those at different times. Could be, you know, sort of a gamer society at one moment and then something that is, you know, related to academia and the next moment or, you know, family management and the next, whatever it is, we're already in these multiple online societies, much more than one. You know, then we participate in the physical space that we're in. So the next step beyond that is, well, if we're already in these other spaces, then transforming that to the physical space that we actually want to be. And you know, how do, how do we think about, you know, forming actual real world corollaries to where we're already spending our mental and, you know, time online today. And so that's the sort of transformation, I think of it largely from the view of, you know, we're creating a new kind of global hyper personalized bank that is run entirely by AIs, is a level of private banking that most people would never have access to and that is enabled by stablecoins, the Internet, AI. You know, all of these things sort of coming together in the middle of this VIN diagram. And that sort of bank is, that is the new hyper personalized bank for people who are forming like physically in their space, a version of this society that they're already joining and already have joined online. It.
- [44:53]
Balaji Srinivasan
Interesting. Yes. And so let's say you, what is the Sha Neville? You know, people ask what kind of company would you start? And you have started a company which is, you know, Katina, and you did circle before that. But I have a different question which is, and actually we also started a currency which is, we started usdc. So if we started companies, we started currencies. I have a, I have a question on the third kind of thing, which is not an Internet company. Or an Internet currency, but an Internet community. If Sean was starting a community, if you're starting, you know, something, what would that community be? What would the theme of it? Is there something you'd like to see in the world? For example, Latin immersion. Right. Or, you know, keto or something like that? Right, Keto. Keto kosher. Okay. There's just examples, toy examples I've got. But you must have something you'd like to see in the world. What would be your ideal community if you could fork Sean and clone yourself and build something.
- [45:44]
Sean Neville
Something, Yeah. I mean, for me, it's about hyper personalization, transparency and trust that doesn't require relying on other human beings. And it's all coded. So one of the things that drew me to cryptocurrency in the first place and bitcoin specifically in 2011, and it was this idea that I don't really need to rely on fallible humans or the. The governance structures that they create that are sort of enforced by courts if you can code those things and encode those things cryptographically in software. And so trust enforced by software is just a fundamental theme. It's one of the reasons I have such optimism about. Although I think you're absolutely right, we're in a little bit of a trough of disillusionment in terms of people's views of. And this switched because nine months ago.
- [46:38]
Balaji Srinivasan
People were really at least beyond AI.
- [46:42]
Sean Neville
But you could mention stablecoins to them because especially in AI engineer communities.
- [46:46]
Balaji Srinivasan
Yeah, yeah.
- [46:48]
Sean Neville
Crypto is even more toxic than funny.
- [46:51]
Balaji Srinivasan
You know what it is actually, it's funny Savior. Say. I was going to say. Go ahead. I was give a meme. You know, finish what you're saying. I was say something.
- [46:58]
Sean Neville
Well, I was just going to say it used to be that we would bend over backwards. Not to mention the fact that our agents were using stablecoins because, you know, people and I think of Circle in a good way is probably the most boring company in crypto because it's a dollar, it's a dollar, it's a dollar. We went into the front door, we know all these sorts of things. And yet people would be afraid to shake my hand because I was a crypto, maybe a Ponzi schemer or something. And it's completely flipped the other way now, where people have a lot of interest in stablecoins, which over the last several months, because of the genius act and other things, have really achieved escape velocity in the mainstream. I mean, it was inconceivable six, seven years ago that stripe would create a payment chain and the VISA would be leaning in, so on and so forth. But it has flipped over the last few months where now people have a lot of interest in hearing about various forms of usage of stablecoins, not even just as a currency, but as programmable money and as sort of a platform to build on. They really don't want to hear about the AI component as much. We're right in the middle of both of those things. But it has absolutely flipped the other direction.
- [48:01]
Balaji Srinivasan
It's like, first it's AI and then crypto, and now it's like stable coins and a little bit. Yeah, you know, it's just funny. You have to sit. Yeah, but it's, there's. There's one thing about that, you know, the meme which. With the guy getting, you know, hanged and he says, like first time to their guy, you know, like from. It's. It's. I'll put them on screen. But so I thought about that because, you know, right around the time when AI was really going vertical in late 2022 and FTX was happening, right, so many people were like, AI is the real innovation. The crypto stuff is all bogus, blah, blah. So many people were saying something like that, right? And you know, of course, AI is a real innovation. But what's interesting today, and huge, obviously, but what's interesting today is, I think in the fullness of time, we see a couple of things. First is AI is actually for many people, in some contexts, equivalent to fake AI scam. A. Is that AI or is it real? You'll often hear people say that, right? So it actually has, in some ways the same. It got pulled into some of the same issues that crypto has. Where is it? A crypto scam is an AI scam, is it? Right. So a lot of the AI guys, like lots of artists and so on, are super mad about AI. And so, okay, that's one piece. The other thing that's interesting is in many ways, crypto is what AI can't do to. That's a deep point. Like I say, because AI is probabilistic and crypto is deterministic. AI can solve partial differential equations, but it can't solve cryptographic equations. And so because it cannot, you know, invert a hash function, a cryptographic hash function, there's certain kinds of NP hard problems, obviously it can't. It literally can't solve them without us changing what we know about computer science. You know, develop maybe P equals np, but we don't know that. So without Breakthroughs in theoretical computer science like AI cannot solve cryptographic equations. And so that, that means crypto's like a hard wall that can constrain and bound AI. Right. It can say, I have a dollar and the other AI or the other human can say, okay, send it to me on chain. And if it can't do that, then it's actually fake. It's just emitting words. Crypto is the actions and AI is the words in many ways. And the actions speak louder than words. Right, right. So in a very deep sense, crypto and AI are complimentary technologies that, you know, you like what you, what you're doing is one example of, I think, a useful synthesis of them with an AI agent that spends crypto. Right. But I think there's others as well. When you start thinking them as sort of peanut butter and jelly, like dual technologies. Dual to each other. Let me know your thoughts.
- [50:40]
Sean Neville
Yeah, no, I think that's right on. And you know, the exact form that it takes. No one. I, I, I, this is like a space where I think I heard somebody mention it's really difficult to play chess strategically in AI because you just have to kind of everything changes so fast. So the thing is to just keep marching the pawns down the board right now. And so people are trying different versions of that, but there are great companies that are experimenting with merging the two in ways that solve real world problems. There are many companies, large and small startups in Cummins that have sort of realized, yeah, these, these two do fit. You know, there is a little bit of a, there's the crypto X AI piece which I think has been a little, maybe a little overhyped. But there's reality there in combining these two technologies at the same time. So.
- [51:27]
Balaji Srinivasan
That's right. In a sense, you know, this is another kind of parallel financial system which is, it's the machine economy that is booting up right now. Yeah, yeah, right. So it's like another whole plane of where things are, you know, trading back and forth.
- [51:45]
Sean Neville
Absolutely. I mean, I think ultimately, Balzi, I think it's going to change everything. I mean, I know, you know, the closest and the kind of analogies that we use are, well, there are so many, you know, people who first experience the Internet on their devices. They sort of skip the whole laptop era and maybe there'll be other people who suddenly begin to experience the Internet only through, you know, agentic interfaces. And I don't think that's a great analogy because I think the change is even bigger. Are we still using Web browsers to interact with the Internet. If our interface is always going to be some series of agents. We're not trying to crack the UX problem here, but these are the kind of questions, you know, that emerge.
- [52:22]
Balaji Srinivasan
I mean, REPLIT is genuinely replit and chatgpt and Codex and Claude code are genuinely new interfaces for. I mean, because they can accept probabilistic input. It like a lot of it, right? As opposed to, you know, it's the exact opposite of how you type into a search engine. A search engine you sort of invert and you find the least frequent keywords in your head and you give the fewest characters and boom, you've got a search query and with an agent you just like you write a huge amount of text and you're as detailed as possible and, or necessarily a huge amount, but like you can write quite a lot and it'll take all of that and your vocabulary actually constrains the machine. And it's just a completely different, you know, interface in essence. I mean, one other thought I had for you, which is we didn't discuss this too much yet, but there's a lot more robots of different form factors, not just humanoid robots, but like robot dogs and drones and all kinds of things coming out of China that, you know, robot locks and so on and so forth, right. That is another potential area for machine to machine. For example, smart locks. Like here's an example, there's public, you could have, for example in Singapore there are these shipping containers that have gyms, like fitness centers there just in a shipping container. Okay. So you could in theory just go up to it and go zip, zip, like this and do a machine to machine payment and open the smart lock dock and have it debited. Right. That seems to me to be a pretty good application. I mean you can do it with Apple pay, but it's lower cost if you do it with machine to machine and it's cool, right? And, and you might be able to do things like that where you know, you've got a robot dog and it's sitting down and then you go like this and it jumps up and it starts running around and doing something. Right. I feel you might be able to do some cool demos there and let me know, let me know if you have any.
- [54:20]
Sean Neville
Yeah, yeah, yeah, yeah. That's a really cool idea. I mean transforming this to the physical is just, is a great space, creative space to be thinking about. Regardless. I do think in those kinds of examples, like you said, yes, you could do this with Apple, but I Think it's going to be very important not to have vendor lock in, in the next version of this sort of agentic web. And you know, I think that's one of the, obviously is one of the things that kind of the web was an amazing sort of culmination of many partnerships and technologies and so on, but it also steered and in certain direction that was not the healthiest and you.
- [54:52]
Balaji Srinivasan
Know, did not do business centralization and vendors.
- [54:54]
Sean Neville
Exactly. And like five companies that matter in terms of controlling identity. You know, for instance. And so the next version of this when we are using agentic interfaces, it ought to be much more broadly distributed and that's the way that it becomes more prosperous for more people and sort of avoiding this, you know, yes, you could. I trust Apple. I don't want to have to trust. And I, you know, obviously respect Apple people. I don't want to have to trust them to get access to that gym.
- [55:21]
Balaji Srinivasan
Yeah. And the thing was interesting is in some ways the society where everything is computational trust actually then loops around and becomes a high trust society again in some circumstances because you know, that person can't defraud you if you checked it. Right. So that means they don't have an incentive to try, even try it. You know what I mean? Right. So, so anyway, I, I think there's something to that which is interesting. Like if that their guy knows he can't get one over on somebody because they'll check it, he doesn't even try. You know, there's like a superior force which is, you know, the, the blockchain like above them that is enforcing rule of code, you know, between them, you know.
- [56:02]
Sean Neville
Yeah.
- [56:05]
Balaji Srinivasan
So. Okay, cool. I, I enjoyed this. And if there's, I guess people can go to Katina Labs right now, now it's join the team and contain labs dot com. Anything else you want to.
- [56:17]
Sean Neville
Yeah, I would say we've been pretty open about some of the foundational sort of agent identity and policy rules efforts that we're contributing just to open source code. Been pretty quiet about the commercial product offering. And so we're hard at work building. We're really excited about it. We're taking a big swing, leverages all these technologies. And so I would say people are interested in from, you know, a partnership perspective, learning more, working with us, potentially. Happy to have conversations.
- [56:45]
Balaji Srinivasan
Awesome. Great. All right, thank you very much, Sean, and thanks for pleasure. Great.