The Agent Economy: Why AI Needs Crypto

Transcript

All right, so I will say I have both been on this panel 3 years in a row probably with Ilya. And this is the first time I think I'm actually positive on the answer to this question. Uh, but uh, you know, we have uh two great panelists. Uh, they both have very different but complimentary views I imagine on AI but we'll find out a little more. Um, so maybe, you know, you guys both have long stories in crypto, have worked at different entities, you know, starting a protocol, creating a a roll up at a large exchange.

But you know, not everyone knows what your your views on crypto and AI AI are, what you're sort of how you've kind of interacted with the fields, how you've been part of it. So maybe give a little background uh about, you know, where you are with your crypto and AI journey. Ilya's begins before he started NEAR and probably so. >> Yeah, so I mean, obviously my AI before crypto uh has been working on the transformer architecture at Google research and then starting NEAR as the AI company uh back in 2017. Uh, we were a little bit too early uh to effectively write coding.

And then realizing that we actually need a blockchain protocol so we can coordinate people, data, and over time more things in a decentralized kind of globally accessible way. And so started NEAR protocol in 2018. But over past, I guess, like 2 years been focused on on how do we A make sure that AI stays private, verifiable, and, you know, anti-swarm uh uh detection and anything like that. And at the same time, how do we actually leverage kind of blockchain markets economy to really unlock I call it right now AI as a single-player mode. And so, how do we unlock this multiplayer mode by using

blockchain as a kind of back end to AI infrastructure? >> Yeah, and I'm Jesse, a less illustrious AI past than Ilya, but yeah, I've gotten to watch Base grow as a kind of platform for AI agents agents tokenizing themselves, buying inference, genetic payments. It's been cool to see that develop over the last couple years. And then personally a deep connoisseur of being able to write code again after 5 years of not writing code and just getting to plug back into the matrix.

Which has honestly been one of my favorite things about this year where it feels like we totally crossed the chasm where, you know, a year ago I was trying to write code it was still this like starting stop thing and today it's like all day every day I'm writing code. In between meetings and that experience I think has been I think one that many of us have been through and we're reaching this convergence point where something's going to happen. And it's happening, but we're we're not quite sure exactly how it plays out and my bet is that AI and crypto are going to come together in some really profound way to make that thing happen and and make it a better thing for the world.

So, I'm excited. >> Yeah, actually I I remember and I met in 2018 and he was like, I have a program synthesis company, which is what Vibe Coding was called prior, you know, in the the prior dark ages. So, you know, you both talked a little bit about AI and crypto and and sort of different ways you think there might be symbiosis. I think historically people have been sort of pessimistic, maybe didn't think there was, you know, I've I've been on the >> not people you Hey, hey, hey. >> [laughter] >> Not just me, though. I've been on like I said, you and I have been doing this panel since 2023. And I don't think people were universally positive, whereas now I think the vibes have changed a lot. So, I guess let's just

start with like what what are you most optimistic about the convergence and convergence and symbiosis of AI and crypto? >> So, I think there there's things that I hear right now, right? Which is how do we make sure we have kind of universal access to AI? How do we make sure when you use AI, it's actually the thing you want to use, not somebody inserting like malicious tool outputs into your agent? Like that's available right now, right? We have Near AI, we have a bunch of other providers. We have obviously Venice and kind of few other places where you can now tap into. And you can pay with Zcash, you can pay with Bitcoin, you can pay with kind of any crypto. And I think like that's kind of now universally like here.

The cool thing I'm really looking forward for the next year is And I And I And I use this example because like people like, "Oh, you know, your agent will be able to do your groceries." And it's like the reality is the consumer apps are so well optimized that like the the change in experience is not that much. What really is a suboptimal experience is all of the business commerce. Is the finding suppliers, doing billing and invoicing, making sure legal contracts are redlined because you actually have a very trust kind of lack of trust environment in business transactions. And this is where I think AI and blockchain really coming in together. Blockchain provides us this trust infrastructure, escrow, payments,

derivatives, all of the things we can do. And AI does all of the kind of emailing, heavy lifting to do actual discovery of products, and do some facilitation. And so, I think that's actually when we see, you know, trillions of actual economy moving on chain, not just speculative, not just kind of assets, but actually doing real commerce work, you know, shipping, you know, steel and doing like chemical supply chain. All of that things starting to really kind of move to blockchain as this AI and crypto coming in.

>> Yeah, I think that's going to happen. I also do think it's going to come on the consumer side, particularly consumer finance intersecting with AI. I mean, I had the personal experience where I saw someone just a few weeks ago tweet about how they were using something called bean count, which is like open source ledger. And I was like plug in all of my I said this to my agent, plug in all of my accounts, put it into this ledger, and then like give me my books and my P&L and everything that's happening. And the agent was able to just do it. And that was kind of a zero for one moment where I didn't have any of that visibility into my personal finances, and then I did. And I think money in general for so many people is this confusing thing where they maybe

don't know how to invest, they don't know how to budget, and it's often felt like oh, I need to learn and it's going to require reading books or going to school or whatever it is. And now we have these insanely powerful agents that I think we can plug into our financial lives that are going to be able to manage our money, trade on our behalf, and really simplify how we all interface with money. And so for me, that's this just super exciting thing that I feel like we're still just scratching the surface of. So I'd say that's my take on the kind of like consumer utility of AI and agent and crypto. Now, on on the maybe more philosophical side, I do feel like you know, you have this kind of like

massive explosion of of content being generated with AI and like what's real. And there is this convergence with crypto where crypto provides the hardness. It's like the verification layer. It's the infrastructure that feels almost perfectly suited to kind of counterbalance the explosion of everything that's happening with AI. And so, again, it feels like they're converging in a way that's going to be pretty profound in terms of how our world is shaped over the next couple years. >> Yeah, I would say that generally when you talk to any people who working in AI, they usually have like this like single player mode. And we've been trained in this like adversarial multiplayer, you know, kind of aligning incentives. And it feels like, yeah,

bringing these things two together, right? The intelligence of AI and the multiplayer adversarial together. And you've been writing a lot about just like applying the same primitives to AI inference markets and kind of just generally compute markets. I think that that's like the missing piece of in many ways for AI thinking. >> I mean, yeah, to your point, a lot of stuff that's been coming out sounds like crypto. So, like a lot of stuff on the consumer app side, they're just been things about like security protocols for consumer agents. And you read the the docs and it's like, wait, you guys just made a pseudo blockchain. I don't like what you're talking about the being like I think AI people actually rediscovering blockchain stuff a lot and they don't realize it because they hate the word blockchain.

It's like a bad It's like a, you know, bad word for them. But I think, you know, if let's take a little look back at the last few years of of crypto and AI. I think we had a few phases, right? The first phases were focused on rent a GPU, like trying to kind of like a little too skeuomorphically to Bitcoin be like, I can get a node. Here's a node for you. We have a decentralized network of nodes. One will be available. The interesting thing about that was crypto people got the market right. There's this huge market for inference providers that, you know, is a tens of to hundreds of billions dollars a year market right now.

It's just that somehow that adding a token to it didn't make it a better product, right? Uh after that we fell out in like decentralized training or AI agents. In fact, AI trading agents were probably just too early, like the AI XBT and and things like that. So, what do you think if we look back that crypto founders got wrong, you know, because like it seems like they've actually been kind of early at the right idea, but they kind of got the wrong framing or the wrong go-to-market and and and what sort of advice from your perspective would you give kind of founders to?

>> I think what what happened is the crypto people who actually wanted to do AI properly just went and did AI company, right? You have Alex who went and did Open Router. And uh he's like, "Okay, we don't need crypto. We can accept crypto, but we don't need crypto underneath, right?" And as you said, right, they used the same principles. Same as like all the miners, like Corvive is effectively ex-Bitcoin miners, right? And like now, you know, half of the neo clouds are effectively miners. Um and so you kind of see that across like few different uh like parts of the stack where when people like, "Okay, we need to go do it." They kind of like, "Okay, all the blockchain stuff is kind of

slowing us down and like really limiting." And I think there is few reasons conceptually for that, right? One is where we normally say like, "Oh, you know, the token is a great incentive to like bootstrap network effects and do this." I would say like past 2 years that hasn't been really the case. I think the kind of new token meta hasn't really played out as well. And then the reality is like token is a product on itself, and so you're kind of like trying to build token and actually AI product at the same time. And, you know, building good products in AI is already hard. So, I I think it's also just like split at the focus. Um I think now that

like some of the companies may actually like use a token when they already established their product and they can use it to bootstrap the network effects as well as the market I would say turning. Like I think some of this is just like literally the timing where, you know, like the AI AI 16Z and things like that like launching a token first and then trying to build a product, like that's a really bad, you know, sequencing, right? We which is they could have built effectively like product first and then launch a token to like bootstrap the the economy. So I think that that's, I mean, kind of historically been one of the challenges.

>> Yeah, and I think that last point you made is almost like a universal rule. That's very hard to um have a early-stage product and launch a token very early and make that work just because there's so much iteration that you have to do on the product and the second you have to also manage this economic system. I think it just adds another layer of hard mode. There's some folks who can do it like I I look at like Venice. They launched it relatively early in the life cycle, but you have Eric leading Venice and like he's he's had a decade of experience navigating those waters.

I I do think that like I'm not even sure that's necessarily that people got things wrong or that like the past eras like they're mistakes. I think a lot of it's just timing. Like these things take a really long time and the crypto systems if you look at how much they've matured in the last 5 years where we now have stable coins, we now have tokenization of equities and all these other assets. Um the chains are fast and cheap and easy to use. Like that infrastructure didn't exist 4 or 5 years ago. And the the other side in the AI side like the models weren't at the point where um you know, I think they were mature enough to really make up for the crypto systems. And so, you know, could crypto people maybe have invented OpenAI or Anthropic? Sure, but I think like my

guess is that again they're just going to converge and that the crypto stuff will have matured, the AI stuff will have matured and over the next few years um we'll start to see more and more of it coming together. And it's happening, right? Like you look at Venice and it's like it's private AI that's a consumer experience that has a token built into it that's working. It's working for the business, it's working for the the asset, um and I think it's working from a agent and AI perspective. And so, um I'm pretty optimistic that that that kind of flywheel is going to keep spinning and people are going to keep finding those intersections.

>> Yeah, so I think you bring up a a great point with talking about some of the successes that we've seen, um especially recently this year, right? So, I think a product like Venice is also sort of a kind of a consequence of open models becoming better. And the sort of open model ecosystem is a lot closer in ethos to crypto and sort of the genesis of it looks much more similar. So, what do you think people have gotten right? And like why has it flipped so much in the last 12 months where crypto and AI projects have been able to get a lot more mainstream success, um you know, everyone from Noose with Hermes, Near, Venice, etc. Some had tokens, some didn't have tokens, but they all kind of

start from the same place. What what were the biggest drivers of success do you think for for these projects? And sort of where do you see them going for the next 12 months? So so next time Ilya and I are on this panel, we have a new answer. >> Great product. I don't know, that seems like the consistent theme. I've had my Hermes agent for like 6 months and I love it. I use Venice and it's a great product. Near clearly has found product market fit on the intent side. And so, again, like crypto as a economic model and as a incentive tool and as a way of capitalizing businesses, um I think obviously a superpower, but you still need the great products. And I feel like generally we've turned the corner as an industry where that's what we're

laser-focused on. That's what everyone's being evaluated on and that means that unsurprisingly we're starting to put up more wins. >> Yeah, I agree. I think focusing on the product and then at the same time the catching the wave of effectively I think what happened that open weight models gotten way better, right? Way closer to frontier from a distance, but also just generally the level now, you know, I mean the example we use is like we run deep seek flash for formal verification and we beat everybody else, right? Including GPT soul, you know, Claude uh and like custom models that people with hundreds

of millions of dollars trained, right? So like the models are actually really good, they're cheap and you can run them for longer because of this. And so you can run agents, you can run, you know, kind of private compute and you're actually getting uh real results now. And so I think like that was a big movement. Obviously Hermes caught on kind of this personalized agent movement as well. So I think I mean as usually with any startup, right? Like the category definition is, you know, you you build a great product and you caught a kind of some wave that you're able to capitalize by being there and really providing that great product to the expanding audience.

>> Yeah, I think one thing though is that the crypto founders have been able to do a lot better with kind of being able to mesh together different aspects of of AI products. Whereas like I think maybe a year ago the open source ecosystem wasn't very good. The kind of ability to to go from consumer product down all the way to like model that's performance, harness that's performance, etc. wasn't available. So I think it looking a bit forward, you know, I think as these models have gotten more ingrained in people's lives, people have started valuing the privacy aspect a lot more, right? And so that's where you're starting to see a lot more of the overlap with crypto. So I'm curious, how do you see pri- the demand for private

compute, private inference, private AI evolving and how where does that intersect with crypto? Because both of you have sort of brought up the privacy aspect in different ways and are sort of more closely involved with that. So so how do you see that evolving? >> Yeah, I mean I I would say like so we launched our private offering December last year and kind of when we started like offering and selling it, the the market wasn't ready, right? A lot of people were using frontier models. So you go to businesses they're like, yeah, we're fine, we're experimenting, it's not critical.

And I think even last three, four months effectively the page has completely turned. People are like, okay, well, I'm actually feeding all of my company data into these AI models and so anyone who has access to these logs can reconstruct literally the company from it. It is making critical decisions, so I do want to know what exactly is happening, right? I want a way to kind of inspect and and like know the exact model, right? not you know, Astra first week was great, second week became you know, way dumber, right? So like I don't want that to happen. And then from even consumer perspective, right?

People are like, oh, actually I'm also feeding more and more of my financial data, my medical data, my all of these decisions are now there. So it's it's also important that, you know, this is not going somewhere. And obviously every single time there's a open AI news that, you know, they potentially trained on some uh you know, data that they were not supposed to and effectively used it to to neighbor stocks or uh kind of example with Anthropic launching Claude design after being a Figma partner. Like every single like that is not helping and obviously them coming out and saying they're going to kill, you know, there's 10% chance everyone. So like they're kind of also undermining their own

position by doing that and so people want more and more like, hey, we actually want to be in control." And so, and that is literally what crypto been kind of like from a philosophical perspective, "Hey, this is sovereign. This is control. You You get kind of privacy, verifiability, and access." And so, yeah, I think it's been massive transition to this just like across all of the things that are happening in the market. >> Yeah, and I think we've seen that firsthand internally at Coinbase actually, where um if you look at the trajectory of our usage, you know, Coinbase is a big company. I think we we push pretty far out on the kind of like overall AI adoption. I'd say at the beginning of the year, um we were all in on the frontier models. Um we were

pretty locked in on one provider that basically everyone was using. And then actually a credit to Brian, I think kind of end of Q1 of this year, he basically said, "We need to not be locked into one model or one one provider, and we need to figure out how to use open-source models, and we need to get into a place where we control our data." And over the last 6 months, what's happened is we've basically massively diversified, where now we run our own harness, we are routing across some frontier models, but then a ton of open-source models that are available to us, and we're being much more conscious about where all of the data flows in the business. And if you look at overall token consumption, it's like up into the right. But if you look at costs, we've basically been able to kind of like get it to tail off and

flatten. And I think from a productivity perspective, we've made the team way, way, way more productive and and successful with AI than we were 6 months or 9 months ago. And so, my sense is that that same trajectory will probably come to every business. It will come for open-source. It will come for privacy. It will come for control and cost management. I think it will also come on consumers, but they'll probably be a little bit further behind. Although it's obviously happening to some extent, but yeah, like people want sovereignty. That's why we're all in it. I think crypto people are a little bit earlier on that curve, but I expect it to go everywhere. And one thing I wanted to just kind of separate from the privacy point, but I wanted to react to as you said this thing about crypto founders

are a kind of uniquely good at figuring out how to work across the stack. One of the things that I'm most excited about is is kind of like a part of Venice's product stack, which is what they've done with DM, where they basically said, "Hey, we have this consumer product, but let's figure out how do we like literally tokenize the inference and put it on chain and then start building a market around it. And I think that's such a great example of figuring out, "Okay, we built a great product. Now, how do we leverage the crypto thing to make it so that product can be more like a platform, so that can be more composable, and so that honestly we can start to invent new markets that increase access, give people more control over their underlying inference, and again, increase sovereignty of the way we're all using AI." And so, my hope

for how we kind of see crypto and AI evolve over the next few years is that we are the people who are like steadfast fighting the fight to be like, "Yeah, we're not going to just have a few labs controlling the future. We're going to have open economies. We're going to have open source. We're going to have open systems." And that's going to be the future we build on top of this new platform. >> Yeah, so actually I'll get back [clears throat] a little bit to the inference market side and how that's that's kind of grown. But actually, you made me think about one question that you know, there's a lot of things crypto people were a little too early to. One thing about kind of the Web3 movement in you know, 2019 through 2022, there's a

big focus on like owning your own data, right? Like you get compensated for your own data, anything you're kind of providing kind of gets gets at that. In the current AI world, every frontier lab, every open even the open source companies are mining your traces to improve their model, to improve the harnesses, to improve And there's no real actual compensation. In some sense, a lot of the original web 3 ideas about data ownership, data provenance, and sovereignty around that should apply here. So, do you think that will be that will happen in the in the next few years, or do you think it's it's still too far out? Crypto people were just a

decade early versus a few years early. >> So, we actually have a product in progress. We got we got to tease it, but it's effectively you I mean, nearly I provide you private inference, but if you do if you want, you can actually contribute your traces back and get paid either in more inference or in in some other way. Uh and effectively, that data would be then used to uh train models. And so, the interesting thing is actually like figuring out how to value that data. Because one of the problems I think kind of when we were talking about, you know, own your data before has been we haven't really like had a good way to evaluate how much

like how valuable this data is, right? And it's like it's just a bunch of random, you know, articles or bunch of, [snorts] you know, documents. This can be anything from worth nothing to, you know, this is a internal Coinbase documents and like highly valuable, but also shouldn't be uh shared. And so, I think [clears throat] now with kind of agentic traces, there's is a lot clearer way to evaluate this because you effectively have human input to like course correct the model. And that is extremely valuable and and and can be evaluated uh kind of programmatically. And so, uh so I think like we'll we'll see this, but it needs to be combined with effectively verifiable inference, so you know that

the AI output is coming from indeed the the So you kind of needed like to build the stack up to get to this level of of kind of contribute like data evaluation and and uh like valuing it. >> I think it's I think the costs are coming for the labs. I'm not sure it's going to start in the consumer side. One interesting trend that we're seeing is that we we've worked with AWS and Cloudflare who both independently um have basically said, "Hey, we have all of these businesses that are running web businesses where they're being incessantly scraped by labs and agents and they're losing monetization as a

result of it and we don't know what to do about it." And the solution that they've come back with is they've said, "Hey, let's empower these businesses to charge on a per page basis for that content." And so we've worked with them to basically make it so like if you run a website on Cloudflare AWS, you can check a box, you can set per page pricing. It can be like 1 1 millionth of a cent per load. It gets charged with X4R2, you settle in USDC on base, and now there's a business model. So that news site can charge a 1 1 thousandth or 1 1 millionth of a cent and it adds up because the agents are mining it or the labs are mining it. My sense is that what what may happen is that over the next year you basically see this shift

where the labs have been able to get a lot of free data from websites and the internet over the last 5 to 10 years and all of those websites are basically going to get hip to it and be like, "No, no, no, no, no. Like you're going to pay us for our data." And then at the same time you'll basically see that kind of increase the value of the underlying data that's being fed into it. And then consumers are going to realize where it's like, "Oh, if all these businesses are monetizing and this is a narrative about how they're monetizing, that's buoying their stock stock prices, why not me?" And then we'll see the exact same thing happen on the consumer side.

I don't see any reason why there shouldn't be like the products you guys are building where it's like, yeah, you can trade on my data, but you're going to pay me for it. Like I, you know, I'll Yeah. Yeah. I think there's a ton of opportunity there. >> Well, I think it's interesting both of you highlighted something funny, which is, you know, I think about the Web3 era, I think about 2021. It was really focused on like consumer personal stuff, right? Like the internet for the consumer that has sovereignty over their data and life. But in both of your examples, you focused on businesses sovereignty. Like the businesses care more about valuing their sovereignty than consumers. Which is kind of a evolution of Web3 in some ways to this market, which I think is is

something that's made the crypto aspects of the current AI market much more amenable. >> And I think it's just indicative of where we are as an industry. Like if I look at what we're seeing from a product market perspective, it's basically like trading, payments, and financing have hit mainstream adoption among businesses where it's just like easier, faster, cheaper, and better for them to run their business on this new rail than the old rail. And so that's leading to like businesses moving over. And then the consumer still benefit because they get better products, but it's no longer like I as a consumer need to be a crypto person.

It's like I as a business can choose this better technology to let me run my business better, and then that's going to benefit all of my customers. Which honestly is a pretty cool change. I think it's definitely led maybe to a little turmoil in terms of crypto's like emotional state over the last couple of years. And identity, yeah. And identity cuz I think we all are pretty used to like selling people the crypto thing, but now we're no longer selling the crypto thing, we're just selling better infrastructure to help people build better products. So, I don't know. I think it's cool. Change though.

>> I mean, I I think it's like it we see this with Venice as well, right? There is a consumer story as well. It's just like you really need to build a product that's like actually matches consumer like level of experience, right? I mean, this is what we're doing with near.com. Like it's fully chain abstracted. There's no gas. There's no chains. There's no like kind of complicated things. Like it needs to just work. And when you hit that bar, then like consumer adoption comes and then you can educate them what the value of privacy, what the value of sovereignty is, and how that enables them to do more things.

But I think crypto has struggled to get to that level experience for the most part. And like there's a lot of like infrastructure challenges that were there, etc. With businesses it's easier because as soon as effectively regulation lifted, it's like, "Oh, it's going to be like 100 times cheaper for me to do this, you know, payments thing. Yes, we're going to do it, right?" So like the adoption there is very like calculated, you know, spreadsheet based than uh kind of oh not always a product kind of experience based. >> For sure. But, you know, saying web 3 for businesses does sound a little like corporate slop. [laughter] It's like, you know, kind of sad.

Um >> It's sovereignty for businesses. >> Sovereignty for Yeah, that's a little bit better. >> Yeah, it it might need a little rebranding. Um all right. So, we got 2 minutes left. You know, you both have talked a lot about new marketplaces that have evolved coming out of mixing crypto and and web 3. I think uh I have certainly have been thinking a lot about this lately and and writing about how the open inference marketplaces, as well as sort of harnesses holding the user data, sort of looks a lot like crypto. You're unbundling the frontier labs. The harness stores user data. It's like a wallet. The router finds you where you're supposed to go. The models are

like public goods. And then the the inference writers are like, you know, liquidity providers. So, I think now that we've seen that crypto-like market structures are showing up in AI, where do you think the next year, like the next big new market structure will show up? It, you know, we have inference, we have data, you know, we kind of have like other forms of user monetization. You know, as you both sit in your seats of of of different types of uses, where do you feel like the the next big new marketplace in crypto and AI will come be?

>> Well, well, the the biggest one is the GPUs themselves, because the the market is completely broken effectively. It's full of OTC transactions, right? And it's missing a place to settle. And so, like that that is like the biggest opportunity right now, and there's a lot of people running at it. Um, but I think like the next after that, I I'm really excited is like this just broader commerce supply chain, like really doing kind of business-to-business relationships, and kind of leveraging web 3 and AI in one place to really deliver, like effectively compressing all of the kind of I would say like toxic toxic costs of doing business.

>> I feel like I'm going to be in an Ilya shell, but which I it seems to me like it's something intent thing. Like, I'm not quite sure what it is, but as more and more of all of our interface and work starts to route through harnesses, whether it's an intent to get inference or an intent to get data or an intent to execute a trade or an intent to do commerce, like it really is taking a human idea, packaging it up into some standard format, and then figuring out the best execution for it across a bunch of different providers. And I'm not sure what exactly the shape of that thing is or how the market structure of that marketplace exists. Um, but, you know, I I was just

talking with someone on my team, and it's like it's intents and order books. It's like aggregators and marketplaces. Like you have the place where the supply builds and then you have the place where the demand aggregates. And those two things are going to come together in some format that I expect will be very agentic and very crypto and very very valuable. And so that is the thing that I think is probably going to emerge in the next few years. >> Yeah, there's you know, been a lot of talk of like outcome-based pricing. Instead of paying per token, you pay for what your outcome is and then And then that that that's sort of like the AI language for the same thing. So it's like I think I think this is coming. It's emerging. So with that we're out of time. So thanks

everyone. Hope hope you got something interesting. >> Thanks everybody. Thank you.