Jake Brukhman (CoinFund) on Anthropic Fable 5, Agents & Asymmetry in AI

Speaker A: We're speaking today, two days after the launch of Fable 5, which is the most powerful model to be released to date. A world where a single AI company that has the best model can decide who can see what. You want decentralized AI training because you want a counterbalance to this centralization of power. There's no one entity who can take it offline. Agents are some of the most decentralizing technologies there are. In about a week, the agents have had reconstructed the same algorithm that Google kept private.

Speaker B: It.

Speaker A: They've lasted through Google's kind of Frontier and then they published the whole thing.

Speaker B: The most interesting investors are the ones who are very early and who are comfortable with the uncomfortable. Coin Fund is one such very early crypto funds which was founded back in 2015. And today, especially in this year, as many are, uh, talking about how crypto as an asset class is feeling, we have Jake Brookman on hash today. Jake is one of the co founders of Coin Fund and he disagrees with this. In fact, his thesis is that AI and crypto are converging and this convergence is not just for distributed training of, uh, frontier AI models, but also in the ownership of frontier open models. We get into this on Unhashed. Jake. Welcome to Unhashed.

Speaker A: Uh, hi, Sree, nice to meet you. Thank you for having me.

Speaker B: Jake, uh, uh, you know, how was it back in 2015 to be one of the very first pioneering, uh, uh, crypto funds? An investment fund looking at crypto. We didn't even probably have the term digital assets back then. So I won't even call it a digital asset investment firm. So it was purely a crypto, uh, fund. And I think Pantera was probably there around, uh, a year or two prior to you starting up. How was it back then?

Speaker A: Yeah, um, you know, when I, so I got my First Bitcoin in 2011, um, you know, I had a friend in Brooklyn. It's a famous story that my friend Will showed it to me and sort of forgot about it for seven years. Um, but you know, when I was thinking about starting a digital asset or, you know, cryptocurrency focused investment firm, there wasn't even, um, a well socialized concept of smart contracts in the market. I was looking at the stuff, um, kind of prior to Ethereum's launch. In fact, Coin Fund launched, I think 29 days before Ethereum, uh, mainnet, but the paper for coinfund was written the previous year. Um, and so this was really for me about trying to understand blockchains. And what were blockchains Good for what were they trying to do? And specifically what were the use cases beyond just kind uh, of digital money or digital digital assets, um, that could be implemented with blockchains. And that's been you know now almost an 11 year uh, journey. And I think we've seen a lot of ideas. Some of them have worked, some of them have not. Uh, but it's still actually quite early I think in the world of digital assets.

Speaker B: Yeah, no, absolutely. Uh, but again, take me back there. How was it finding your first few LPs uh, how did the fund uh, come about it?

Speaker A: Extremely difficult. I mean in 2015 there were no professional LPs looking for uh, exposure to the crypto space because the crypto space was just also not a very uh, big or um, um, you know, just well understood space. So our first fund was really about um, friends and family backing us. And then uh, as we continued to raise funds, you know, our you know, that LP base, uh, became more about high net worth individuals, family offices. And then today you know we have some extremely professional um, LPs including sovereign wealth funds and uh, and fund of funds and so forth.

Speaker B: Right, okay. And uh, whilst I was just sort of reading up uh, for unhash, I mean for today's chat with you and uh, you know, realized how early coin fund was, uh, and also a bit of your sort of background, um, you know, popped up there and I couldn't help but uh, sort of reflecting on the fact that this other guy who also sort of was born in Russia and had computer uh, scientist uh parents and uh, ended up uh, building uh, sort of Ethereum. Of course, that's uh, Vitalik. And sort of how was it for you? I believe you were born in St. Petersburg and uh, to computer scientist parents as well. Um, how was all that and when the move happened, do you think that somehow has played a role in how you took a contrarian view to many things, how you approach things in life?

Speaker A: Absolutely. I mean, so I was born in Russia. My family is ah, like mostly Ukrainian Jewish, um, came here when I was about 6 with my parents in 1990, which was just sort of a crazy move uh, for my dad to pick up his family and move to a totally different country, especially out of Soviet Russia at that time. Um, one of the ways I think that has shaped my entrance into this space is I think historically blockchain and crypto has been quite um, philosophically aligned against centralized power, against really powerful oppressive government. It's a check and balance. And so you know I come from kind of a cultural background with you know, a communist government and sort of like a lot of distrust of government and which has taught me to be skeptical. Um, you know, that skepticism I think was very early or very, very necessary in the early times of crypto. Because a lot of people, you know, back in 11, 12, 13, 14, 15, you know, would look at, and even today still, you know, look at Bitcoin and say, you know, this isn't serious, this isn't backed by, this isn't um, a real asset, et cetera. And I think Bitcoin in particular has proven that to be a bad strategy. And the people who were sort of contrarian very early I think did very well on that. Um, my parents were. My dad was actually trained as a physicist. When he came to America he realized he could really make money working uh, in university in New York doing that. Um, and became one of the first computer programmers on Wall street in the 90s. Um, and then kind of transitioned into quant trading and things like that. And then my mom was, ah, she had studied computer engineering, was also working kind of the same field on Wall Street. Um, and so we had a computer in our house from very, very early time. I think we had a Sun workstation or house in, I want to say it was like 1991. Um, so even before most people had kind of even enthusiast Internet. Um, and so I learned how to code C from my dad on our sun workstation when I was about 14 and so on. Um, that kind of kicked off for me being an early adopter of various technologies, computers, the Internet, programming, mobile, um, you know, eventually crypto, eventually self driving cars and so forth. Um, so yeah, I think my parents definitely shaped my view into my career as a technologist.

Speaker B: And um, you mentioned about the origins and the philosophy of um, how crypto spoke to you as a distributed, decentralized, uh, system. Uh, and uh, you know, which, which stood up to the powers of centralization or concentration and which is something which has sort of attracted a lot of us into the whole space, uh, in the first, uh, place. But now, um, more recently we're seeing the old institutionalization of uh, uh, crypto for, for good reasons. And I think uh, as I saw this uh, you know, this whole banter, uh, around how defi, uh, there's no defi anymore, it's on chain finance. And I think you tweeted saying defi is a feature now, right? For now, yeah. Uh, so much has changed in terms of what the original promise of defi was in a way. Uh, um, or do you think that this, this, this was the path, uh, that was uh, ordained or was bound to happen for something once it scales up?

Speaker A: I mean, I think if we rewind back to, you know, kind of the coin fund team sitting back in, I would say like Q1 of 2018, um, you know, this is like right in the wake of the 2017 ICO bubble crashing. It's, you know, we're kind of sitting there watching the US regulators, you know, start to perk up and start to issue like various threatening guidance. Right? Like, you know, our sense was very much that, yes, we want this technology to be ideally, um, adopted bottom up. It's the kind of decentralization vision where everybody gets a wallet, puts a little bit of ether into it, builds some decentralized applications, and then creates a new economy in parallel to the economy that exists. But at that moment in Q1 of 2018, we were also very eyes wide open about the fact that that might not happen. And the way that like, if we want this technology to be adopted, um, and really like, make it big and really be useful and make an impact in the world, it is going to have to negotiate. It's going to have to negotiate with governments, with banks, with incumbents, with enterprises. And it has to um, really create real value props and you know, educate people and um, fit into laws and compliance and things like that. So we were always like, from a very early stage, um, we're a fund who is thinking about the convert. We called it convergence, right? The convergence of this extremely new, innovative, disruptive technology and the real world. And that is how we sort of built our investment discipline. That's how we hired for our investment team. And that, you know, has philosophically been the case now. We've also seen a lot of decentralization experiments happen. I think those experiments are important, they're valuable. You know, the fact that like you could have some semblance of decentralized governance, you know, that's a good thing to know. Also the gaps in, you know, and the, the downsides of having decentralized governance are good to know. Like, I do not regret, you know, seeing a lot of founders, you know, run, run those kinds of experiments. I think now we're at a moment where, um, first of all we have the convergence of three things that really have not happened before, which is, uh, first of all, a favorable regulatory environment for crypto, a really mature set of technologies where we've solved a lot of the key technical problems like, you know, security, privacy, scalability, historically have plagued the Space. Um, and then we have more adoption from institutions and enterprises than ever, uh, in the sense that like CEOs are sitting down and, you know, integrating this technology. So in some sense, blockchain and crypto have never been more successful than they are now. But they are in the process of, you know, kind of finding product market fit in much fewer areas that have been historically kind of hypothesized. Like some of those experiments like NFTs, you know, DAOs, et cetera. Like, you don't see a lot of that in the market today. It just, it wasn't, it didn't have a ton of product market fit. Um, now what makes me feel better about kind of quote unquote going through tradfi, as someone at a dinner put it the other night, you know, they were saying, you know, for crypto to succeed, it has to go through tradfi. Um, what makes me feel better is that I think, like, that is a process that'll actually put a lot more blockchains out there in the world that people will be able to hold Bitcoin in their bank accounts. Ethereum, um, you know, most banks are, uh, at least familiar with some kind of EVM technology. Right. And so in a world where like, there's a lot more people using the tech and holding the assets on blockchains, I think there'll be an opportunity down the road to create new startups, to create new decentralization experiments. Um, and it'll be a lot more feasible, you know, because we'll have this critical mass of actual users and, you know, real money to be made there.

Speaker B: And you think that some of these sort of experiments that have happened in terms of governance, in terms of the old, you know, financial stack, which has been built with, uh, uh, defi. Some, Some of this will make a comeback in a different form. Right now is kind of probably a transitory phase for, uh, DEFI not being looked at as the future of finance, but as a feature of, uh, finance. Do you think it's sort of transitory right now?

Speaker A: Well, I think. Look, I think, um, the market is demanding, you know, technologies, companies and protocols that are finding revenue traction, product market fit, you know, and like real use cases. Right. Um, the set of those things seems to have narrowed dramatically. It's, uh, you know, kind of a broad category of like on chain finance now, and that includes payments, stablecoins, uh, tokenization, which I think are all sort of the things that we had always said, right. Crypto would be great for, but they're kind of the Boring, unsexy, big things, um, however they're happening. Right. Um, another category that, where we've seen a lot of interesting action, you know, of course is prediction, uh, markets and kind of the set of tooling and data around those things. Um, and then finally a much more speculative area in uh, kind of decentralized networks and Web3. But I think also the most high risk reward one is decentralized AI, which we'll talk a little bit about today. Right. So I do think that, Sorry, I'm not sure if I'm answering your question, but I do think um, we have some really interesting opportunities to uh, you know, in those areas and then as we see success there, you know, I think we'll go out from there and there'll be new areas to invest in as well.

Speaker B: Right. And you know, picking on investing. Um, you've been a very sort of uh, if I were to use the frame, very systems oriented investor, you know, looking at crypto investing as kind of network investing, not just deploying capital but actually participating in it, be it managing a network, being a validator, staking, being part of it. How is, I mean a, do shed some um, light on that in terms of what a network uh, investor is and B, also bring it to today and in today's context, ah, as a, you know, digital asset investor is. Has your thinking changed? Uh, uh, you know, especially, and I'll come to that later especially, you know, when you look at uh, how a token is being viewed in the market, what is the role of a token? How uh, you know, what's the life of a token? What is the sort of financialization, uh, uh, around a token which was there earlier versus now. So when you look at a, back then, the thesis of being a network investor versus uh, now how things have evolved, just put them together.

Speaker A: Well, I think when we had just come into the space as Coin fund back in 2015, 2016, we were really trying to make a thesis saying look, this is just a totally new asset class and everything about these companies will be incorporated like into this new asset class and we have to build a vehicle to invest in these assets. So uh, digital assets, tokens. I think what has transpired over a decade is that you know, people realize, okay, well you know, digital assets have some really interesting properties, but equity is a really good tool for incentivizing a team early before a network exists, you know, doing governance, doing uh, you know, making sure investors are compensated in, in the loop, you know, things like that. And so, you know, one of the dimensions that we could talk about is, you know, we went from kind of just pure digital asset or network, you know, asset investing to, you know, the middle 90% of our deals usually have both. Like it has an equity component and a token component of some kind. Um, and then sometimes we have, you know, pure equity companies that are operating in the space and sometimes you have pure digital asset companies because there's some kind of protocol. But like, the middle 90% is usually like a mix of humans and network, um, you know, in terms of being a network investor again, I think in the first half or so of that journey, you know, there are some like, really interesting opportunities. I remember in 2018 we became one of the largest token holders of the live peer protocol. And that's not because we bought a safe note or an equity note or even tokens from the team, but rather we helped the team mine their token supply in the scheme, in the kind of the public scheme that they put out. And I remember sitting with my CTO at the time, Michael Bosworth, and figuring out how to, um, know, solve these like cryptographic puzzles that basically would enable us to uh, you know, to mine the network. And, and then we became as big of a token holder as some of their major investors at the time and then later became also cap, uh, table investors with flashp. But like, you know, it, our entry point into that network was just purely through, uh, you know, decentralized open, uh, protocol. And um, I would say at the time we were one of the few and only funds that were so hands on with being able to do something like that in the networks Again, I think as time moved on, those kinds of things became a little bit less important versus, hey, is our blockchain getting customers or is it getting traction? Is it doing its job? Um, and so, you know, at this point in time when there's a lot more of an onus on that, I think we're seeing a lot of like very traditional looking companies. It looks very much like fintech. Uh, we're seeing a lot of payments companies. Um, we just announced, uh, our investment at Edge Markets, which is a company facilitating payments across prediction markets. Um, and I think that's okay. I think like my criteria for, you know, is this a coin fund shaped investment? Is this under the mandate of our crypto fund? Um, has always been, are they using web3 and blockchain technologies to gain an advantage? And of course these companies are. Some of them are using stablecoins, some of them are superstate. Is, um, tokenizing the stock market among Other things. Um, and so by creating these new innovative products, you know, I think that's, that's actually you know, very, very exciting. But we see less of the sort of crypto native network, native type of opportunities today. And a lot of them are still in the market. Like you know, things like Uniswap, uh, decentralized exchange protocol and many, many, many others. A lot of defi protocols, they're still there and I think there's still a value proposition there. But I think from here we have to see a little bit more institutional and Wall street and financial adoption of on chain finance before they kind of go uh, and engage with decentralized finance.

Speaker B: Yeah, and we're seeing that happen. Right. Even uh, you mentioned uniswap and uniswap v4 actually has um, uh, permissioned sort uh, of uh, uh, pools. Right. And you have to have on chain IDs KYC. So all of that sort of seems to be um, converging. And uh, in this, as more of the newer crypto companies, digital asset companies, um, come, uh, and they will largely. What's your view on how their structure will be? Will there be a token play as well? Are you open to as an investor looking at uh, uh, crypto startups which have both an equity and a token play, or you're veering around to thinking that either they have to be one way or the other. What's sort of the whole thinking and thesis uh, for Yuancoin fund that definitely I love tokens.

Speaker A: I still, even as mostly an equity vc, um firm, still there are opportunities in the market where we can buy tokens and we have done that. We could take a venture view on digital assets of certain networks. And as crypto vcs, it's really, really important that we have the optionality to do that. Um, I think what you see is actually like an interesting mix. You see some protocols like across protocol for example that historically have been very decentralized and token focused. Actually say listen, we're going to buy back the token and convert that to equity ownership for our users. Um, because we think that that is the best way to move forward with this particular business. And then at the same time you see people um, maybe like working more in the, let's say the uh, infrastructure as a service like uh, space and blockchain. Say look, we've done a lot of staking here but despite being an equity company, um, we can see a path of creating a digital asset that's backed by staking revenues or something like that. Right. And creating A product. And you know I think it's, it goes both ways and it's still like a very active space where people are you know, trying different things. I don't think that tokens overall like as a technology are going anywhere. There's a famous talk from Kevin Kelly from like 25 years ago at Ted, right where he's like, you know, once you kind of create a technology it never really goes away. Um, but what is changing is like what are the, you know, what are, what are the use cases for tokens? I don't think like if Ether as the native asset of uh, the Ethereum network is going away, I don't think that Bitcoin as permissionless, uh, global decentralized digital gold is going away. Um, but there's a lot of other models that are sort of shifting and a lot of new digital assets like tokenized equity, tokenized real estate, right that are actually starting to come into the market that have always been in view but are only being implemented now because the regulation is finally uh, arriving. So I think we will continue to monitor um, the opportunities there as a fund. And our thinking has always been we want to invest in the set of assets that captures value for this business and for this decentralized network as well. And ah, so we'll continue trying to figure out what those assets are and

Speaker B: what's, what's your view on where this tokenization vector sort of really makes sense? Where is it really solving a problem by bringing assets on chain onto, on chain finance for example? You know, I mean there's this whole thing about everything getting tokenized but that, that might not be really fixing or addressing a particular uh problem. So and we've seen also various experiments from art to real estate. But, but the real adoption uh, um, and or and scaling is, is, is happening in financial assets. Uh at the, at the end of the day, I mean largely still it's a chunk of us Treasuries, there's some private credit which was have you know having its moment till there's sort of a recent uh, uh head wheels. So what's your thinking?

Speaker A: I think, I think, I think tokens add. Well I think they saw, I think they do solve a problem and they also introduce an innovation. So the like at the basic level the problem that they solve is operational. They make operations with assets a lot easier. This is why you know you have a bunch of banks tokenizing deposits right now. Our uh, company Matter Labs and the ZK Sync network just signed like five regional banks that are now on Kind of like a single blockchain network tokenizing their deposits because they know that settlements, um, between the banks with these kinds of tokenized deposits is just going to be faster, easier, it's going to lower costs, right? It's going to be cheaper because you're not paying, uh, a clearinghouse necessarily, uh, in the case of equities, whatever, but it makes the process cheaper. Um, sure, trading might become cheaper. I mean, trading is already free for most retail. Um, but when you go from one brokerage to the other, there's this very slow, onerous paper process that moves your assets along. If these institutions connect through blockchains, that process is, could be like near instantaneous. That's going to create more and better competition. It'll be better for the, uh, the user. You know, if I can leave my kind of silly brokerage because they raised some fees on me to some other brokerage which has a better fee structure, better service, you know, I'm going to do, I'm going to be able to do that faster, you know, on these, on these technologies. I think that's, that's great for the consumer. So, you know, there's a set of, I mean, you name it, right? Like tokenized deposits, settlement, um, you know, we talk about T plus three, T plus one, T plus zero, um, sending wires, wire fees. Like there's all these, um, there's all these operational issues that will just become better, faster and cheaper under blockchains. But I think the, the much more interesting thing is the innovation that it creates. So like, so why do you want to tokenize all of your assets? Why do you want to like, sort of securitize them? I mean, the basic reason is like liquidity, right? It's like I own a motorcycle and my motorcycle is valuable. But to turn my motorcycle into money, it will take me like months to sell it or something or find a buyer. Same with a house. Um, however, if all of my ownership in these assets is tokenized, if all my stocks are tokenized, I can actually like much more easily go to services and protocols, you know, and financial, uh, pools, right, where I can get liquidity on my assets. So before, you know, I could get a home equity loan on my home, but I can't get my, a motorcycle equity loan, right? But now I can. Or like before I could use money to pay for my coffee, but I can't like borrow money against my, you know, Tesla stock or SpaceX stock to buy a coffee, but now I can, right? So you, you open up like, I think a whole like second order set of like liquidity services around tokenization that I just think, you know, makes the world a lot more capital efficient. And that's a huge innovation.

Speaker B: Yeah, no, no, absolutely. I think we all sort of, all the efficiency and disintermediation that comes with tokenization is sort of fantastic. And it's there to fix settlement, uh, it's there to fix payments. But really, uh, what you talk about in terms of the creation of new markets and market structures through liquidity and bringing utility to tokenize the structures of uh, real world assets is what's really, uh, the innovation. And that sort of innovation probably also needs an enabling regulatory framework, um, for obvious reasons. And where do you think we are on that spectrum? I mean, of course when you take global markets, it's very different. Why I'm from one region or jurisdiction, uh, to another. But let's take the US and then where we are, uh, in the US and where are we likely to get in some of the other important financial markets of the world?

Speaker A: Yeah, I mean, look, I think globally that regulation has been in the process of being clarified. We saw some action in Europe, it's almost two years ago now. Um, in the US we've seen uh, the Genius act kind of clarify the role of stablecoins. And I think that's been very, very bullish because essentially has enabled banks to work on stablecoin infrastructure and payments. And that is a very active space. And it was a space, by the way, that customers are extremely excited about. They're global firms that struggle with the international banking system. It's slow, it's unreliable. There are certain countries where you just can't trust it. Money gets stuck, stolen fees are high. Right. And so these founders were like, oh man, we have this stable coin that we can send to Africa to our counterparty and it settles in minutes. It's just like a game. Changing technology and payments are kind of boring to most people. So they don't understand the joy that this kind of thing brings to companies, uh, around the world. But it's a very serious, you know, very serious area. So then, so that's, that's just from the, from the genius, uh, Genius act and then the US we're kind of waiting patiently, um, for the Clarity act to be passed. And I think that that will give like a lot more clarity on, you know, digital assets. You know, the, the jurisdictions of the sec, the cftc. What is a commodity? What is the security, you know, in terms of digital assets? And really, like, I gotta say, sri, I was like shocked when a few months ago kind of the SEC and CFTC had this joint statement where they kind of preliminarily published their ideas on what that would be. So essentially they published all the clarifying guidance that the crypto space for 10 years was asking the government for. And this was like finally like okay, here it is. And then the market didn't even. Right, it was, it was like this. Oh, it's just something we've been waiting on for for ten years. Um, so you know, I think it's been very, very bullish in the US for the regulatory landscape. I think I'm hoping clarity will, will make it through. Seems, seems uh, logical. And um, and then I think there's going to be like a couple of years of sort of the industry, the uh, institutions, the enterprise is absorbing that and starting to put out real products based on this regulation and seeing how they work. And we're already in the process of that as I mentioned, um, and I would say like a couple of years from now we're likely to see a lot more widespread use of stablecoins, tokenized assets, prediction markets, um, uh, you know, kind of fast payment networks, fast settlement networks. So we're going to see a lot of innovation and finance uh, happen because of this. And by the way, timelines are also going to get compressed by AI helping people implement this stuff.

Speaker B: Yes, and I think that's the right sort of segue to uh, you know, getting in the convergence of uh, Web3 and uh, AI and um, you know, more specifically if um, I've understood uh correctly I think you've been talking about and this convergence has been sort of one of the major themes that many people have been talking about. But you've specifically spoken about um, distributed computing uh, for uh, pre training, uh, LLMs in fact and uh, I mean just put things in context, if you could put things in context, in fact, that you know, pre training is probably the most um, you know, the more difficult heavy lifting part of uh, training foundation, uh, you know, models and really getting distributed um, systems to be able to you know, uh, provide the compute for that. Is that something which will be feasible? Firstly, I mean especially in today's context where we're looking at heavy GPU clusters being used for uh, that all in the centralized. So deconstruct that please.

Speaker A: Yeah. Okay, so maybe we can just take a step back before answering that question. Just give a little bit of setting. Okay. So we live in a world where Obviously companies like OpenAI Anthropic have grown tremendously Fast, you know, they're heading for trillion dollar type IPOs. Right? Um, so the big AI companies by the end of 26 will have spent about $600 billion on AI infrastructure. And mo. The vast majority of that infrastructure of course is going to train models and to provide inference of models. So inference is when you kind of hit the API, call a model, get an answer, or like when you're in chat and you know, you talk to the model, you get an answer. Um, the vast majority of revenue made by the AI world is on this idea of like giving answers, right? Of, of doing inference. And that revenue looks something like 60, 65 billion of ARR, according to an estimate that's now, you know, three weeks old, so it's probably stale, but something like 30 billion ARR coming from anthropic, about 25 to 30 billion coming from OpenAI. You know, some analysts saying anthropic alone could be 100 billion of uh, err, alone, you know, by the end of this year. And then if you look at the kind of growth rates and sort of try to estimate, you're looking at a market in five years where there's 500 billion to $1 trillion of revenue. You know, for AI inference, that is a huge, huge, uh, number. And you know, as this is happening, and by the way, we're speaking today, two days after the launch of Fable 5, which is a model that is the most powerful model, you know, to be released to date, but is also completely censored by Anthropic. And it says, you know, you can't do biology on this thing, you can't do cryptography on this thing. And by the way, if you're trying to be competitive and try to use it to create a pre training pipeline or to compete with us, you can't do that either. And there's a lot of discourse this week about is this really the world that we want? A world where a single AI company that has the best model can decide who can see what, whether corporations get something before individuals or how much is this all going to cost? Um, right. And so that's sort of the background of why decentralized AI, and in particular decentralized AI training is such a huge opportunity. Right? So philosophically you want decentralized AI, uh, training because you want a counterbalance to this centralization of power that's happening de facto, you know, because these companies are growing so powerful because of these products. And even Yann Lecun today was tweeting about, um, an organization that he backed recently I'll put a look for you. Um, but you know, this is an organization that is a consortium of um, it's called Project Tapestry. And it's a consortium of companies that will come together and train frontier models, but ones that are open. And then the question becomes like, well if you want to do that, where are you going to get the compute? Because this whole game of like we have the biggest, best model comes in the back of these giant high end Data center grade GPUs. And these big companies have, you know, essentially bottlenecked the supply of these things. And so what has happened? Know what would I, what, what would happen if I told you SRI and um, and your audience. Let's take these training processes. These are processes that take place in these big data centers. There are some companies that are buying nuclear power plants in order to power these data centers. They have these enormous GPUs, industrial grade machines with hundreds of gigabytes of memory terabytes per second of throughput between them. And uh, what if I told you that I could train the same big frontier models that they trained in these data centers, but now on regular consumer, 80 megabyte per second Internet, on people's gaming devices and MacBook Pros? Most people would say hahaha, that's really funny, that's not possible.

Speaker B: Right?

Speaker A: And that is exactly what four years ago the AI experts told me when I started like looking into some of these um, companies that were thinking about doing this. Um, and our first investment in the space was actually Sam Altman's uh, world project in 2020. And then in 2022 that's the first time I heard the idea of decentralized training from a company called Jensen. That became our portfolio company. And In September of 22 we published a thesis at Coin Fund on decentralized trading. And that thesis remains the same today. Um, except what has changed is that decentralized training has gone from completely infeasible, completely impossible to. We have many now examples of decentralized trainings having taken place. Uh, the trend is that the number of parameters for models that are being trained in this way is going up. It went from 10 billion parameters last year, company called Prime Intellect in our portfolio. Then it went to 40 billion. Then a couple of months ago a company called Templar on Bittensor treated a 72 billion parameter model. Last week a 100 billion parameter model was trained by another company called Macrocosmos. Our company, Pluralis, you know, has figured out how to do these trainings by breaking up the model. Uh, so not A single entity owns the whole weight set. Right? So these are, you know, what's happening here is that we went from impossible to, this is actually a frontier research area of AI. This is advancing the state of the art of AI. This is now a highly published field with papers in all the major, uh, AI conferences in the world, including NeurIPS. Um, you know, and the trend is these things are kind of going up and now we have this fable five week where everyone's realizing, listen, we need to counterbalance the power of companies like Anthropic and OpenAI. But where do we get the compute? And the answer is when people sit down and think about it and they optimize the algorithms. You can create algorithms that train models on commodity hardware in a swarm and for many applications, and certainly for commercially viable applications, this is now very clear. Um, this is feasible and even cheaper and faster.

Speaker B: Okay, so if I were to kind of deconstruct that distributed computing obviously existed. Ah, but using, um, distributed computing for training, uh, frontier models is something which was sort of still being, uh, tested out. And now what you're saying is that, okay, now that's already happening, now it's feasible and it's grown to model with uh, 100 billion, uh, plus uh, parameters as well. Now, uh, once those models are trained and essentially that model is set of uh, weights and biases, essentially, and those are now owned in a distributed structure. And that's how the ownership of the trained model is decentralized.

Speaker A: Right, so this is sort of a key point, right, because to date, um, most models that counterbalance these big proprietary centralized models have been open models. They're mostly been created by Chinese AI labs and Meta. And the issue with open models is that, I mean they're great, but, and they're proliferating, um, so they do like, enable some innovation. But one of the drawbacks of the way that open models are made is that they're not fully transparent. So users get the weights of the model and they also never get the setup, the data, the cleaning pipeline, the hardware set. They don't know how the models are actually made. Um, once the open weight model is out in the open, then no one has any incentive to pay for it. Everyone has the incentive to quantize it, make it cheaper, inference it on their MacBook, locally, on their own computer, whatever, but they don't really want to pay for, uh, inference. And they're cheaper as well. And overall that has been awesome. That's that like most production processes, you know, you know, you're a Startup and you're using AI to teach people language or whatever it is, right? Like you're probably using mostly open Chinese models in the back end. And uh, they're almost as good and they're, and they're 90% cheaper, you know, and this is just, it's great. But one of the issues is that the more we sort of live in this world and the more powerful the fables of the world get, the more expensive it is for these open models to keep up. And it just becomes an economic game of like, okay, well now I have to pay $20 million, now I have to pay $50 million, now I have to pay 100 million, whatever it is, right, to like train these models. But the business model around those models, like, isn't as clear or doesn't exist. And that is actually why you see a lot of the Asian labs now starting to put out closed models and they kind of keep the IP very close. The opportunity for decentralized training is to innovate on both the proprietary model model and the open model model. And the way that it does that, it says we will separate the weights among different, um, kind of participants who help train the model will tokenize them. Right? And that token will give you, um, the rights to the API revenue for that model. So this is a model whose weight set is not open and that enables a business model. It says, well, if you want to get the output of this model, you got to pay the network. And when you pay the network, the revenue is distributed among the owners. And so you have this model that's valuable. It has a business model, but it's also publicly owned. There's no one entity who can like, censor the whole thing. There's no one entity who can take it offline, um, you know, and so forth. So, you know, it's just a kind of a model that lives in between, that is both valuable and public. And I think that's really, really interesting.

Speaker B: All right. And no, that's, that's absolutely. Actually it's, it's fascinating. And when, when is the first such model going to be out in production?

Speaker A: So to break up a model, you know, in this way. So actually like most decentralized trading runs to date have not been about breaking up the model, they have been about breaking up the data. So a lot of the bit tensor runs that you have seen, they've actually produced open weight models, which is fine. You know, still, it's still a great advancement. Um, our company Pluralis is the only, um, the only company that I'M aware of that is working on model parallel training, which is the process by which you actually break up the weights. And so they have uh, they had a run on this in you know, uh, a model parallel run in October of 2025. That was the first one. We're currently right now in the middle of the second run. They've made m major advancements in that network. And I think what we're going to start to see is that these models are going to start becoming bigger and bigger and bigger. Parameter count is one dimension and then whether that's one trend and then the other trend is whether the model uses what kind of hardware does it use, does it use high end GPUs or does it use prosumer devices or does it use like MacBooks and commodity devices? And I think where this is going is heterogeneous devices are starting to be used. The more heterogeneous devices are possible to be used, the more it becomes possible for home users to participate in large scale training processes. These processes are actually cheaper than data center process. Like if I'm a customer of uh, you know, wanting to train a model, like we heard the other day that Kirkland and Ellis, big law firm has put aside $500 million, train their own model, their own proprietary data. So these guys like where are they going to get engineers, where are they going to get compute? How are they going to put all this together? How much is it going to cost for them to train? Well these networks offer a uh, per run less expensive cost because when you train in a swarm and you're training on people's like Mac Studios and 4090s and gaming devices, you don't incur data center M facilities maintenance, you don't incur cooling costs. It actually works out to be cheaper. The only question is can we scale this enough for the application appropriately for the size of the application? And that's the process that's happening right now is the scaling process. It's going on. Um, and then the other, the other trend to be watching is like okay, are they breaking up? Are they producing open models or are they producing like publicly owned models? And I think that's going to be another important distinction because it is the publicly owned models that will have tokens. And I think that that's, you know, if you look at the interest in the financial markets about people wanting to participate in companies like Anthropic OpenAI like there's really like no public market assets yet to buy. Um, and when those companies IPO Right. That's going to be priced in already. And then on the private markets most people can't get into AI and that market is just like over invested anyway and it's a big bubble right now just waiting to collapse. Um, and so my thesis is like the average investor, when they're going to be getting AI access, they're actually much more likely today to get it through some kind of digital asset than to get it on the public market. Um, and that's the case today. And so once you start tokenizing these models and people are like, oh wow, we built this new cool video generation model like our company Bagel did the other day. Um, I really want to promote that one. I want to buy the token and I want to see if people use it and if it ticks off like I'm going to make money just like I did another digital asset context. And so I just think there'd be like a huge demand for tokenized AI models.

Speaker B: Right. It's almost like this long arc from network investing to um, to decentralized intelligence investing where you're owning uh, AI in an actual AI model, not by owning or investing into a company that owns it, but actually the model in its decentralized uh, form. In a tokenized form.

Speaker A: Yep.

Speaker B: What's your sort of um, um, let's take the next. I mean things are changing fast. What's your 5 year, 10 year uh, view on how this whole convergence of AI tokenization, decentralized, um, compute and inferencing. So how is this whole market going to kind of come uh, uh, together? How is it going to look like in the next decade?

Speaker A: I think there's like two areas of intersection with Web3 that I really care about. So we just spoke at length about decentralized training. So just to kind of finish off the thought there, I mentioned this could be a 500 billion to $1 trillion revenue, uh, market in five years in 2031. I mean if decentralized, if these um, public models Capture let's say 10% of training, you know, um, they also capture the commensurate inference that comes with that. So you're, you know, maybe that looks like 50 to $100 billion kind of an opportunity. Um, and there's more and more companies like, like the activity of training is actually like proliferating. It might not be obvious but even though there's more centralization, there's still more models than ever and more continued uh, pre training processes, more specialization of these models, more et cetera, et cetera, uh, and People care about costs, so always train. There's more and more training going on, so you're looking at just like a huge financial opportunity if you could capture it that way. The other, uh, intersection with Web3 that we didn't talk as much about is agentic commerce, essentially, or how do people use agents and how do they interact. Now, I've given a number of talks this year at East Denver and Consensus Miami about what does Web3 do for agents. Um, I think a lot of Web3 people want agents to use stablecoins. I think that's a candidate. I'm not sure that, you know, necessarily stablecoins will win just because. Right. But where, you know, what I will say is that I think agents are some of the most decentralizing technologies there are. And what actually gets me really excited about agents is not the fact that they're paying each other in stablecoins, but that they're coming together in swarms and they're starting to solve problems together. Together. Right. So we had this, um, really, really interesting, um, thing happen where just two weeks ago, Google published this quantum paper. It said, look, we're going to start to break ECDSA encryption. That's the encryption underlying Bitcoin and Ethereum. Um, you know, crypto guys, watch out. Here's, you know, we made this progress in this algorithm. It's now, it now requires many fewer cubits to work. But guess what? We're not going to actually publish the research. It's too dangerous. We're going to censor it. And so about a week later, a swarm of agents came together, as put together by Eigen Labs. And these agents took that problem and sort of incentivized, not, uh, incentivized, but for science, kind of brought these agents together to try to solve that problem. And in about a week, the agents had reconstructed the same algorithm that Google kept private. And then in another week, they improved it by over 40%. Right. So they just, they blasted through Google's kind of frontier and then they published the whole thing. Right. And so you can really see, you know, this is one case where, you know, you're kind of making an optimization. You can see other cases where agents are solving important math problems. Um, there's like Millennium price problems that are each worth a billion bucks. So you can imagine, you know, someone clever might come up with a math agent to solve that at some point. Um, you know, there's, there's algorithms that, that need to be optimized there. Like, there's all kinds of stuff that these agents could come together. And there's a, and there's a few startups that are creating platforms for the agents to do that. Right. And you know, all of that is decentralization technology. It's like coordinating agents with a decentralized memory so they can work on the same problem. It's giving them financial incentives like, hey, if you solve this, here's a tokenized kind of um, ticket where you're entitled to the prize, if we ever get the prize. And so I think there's just going to be a ton of use cases for Web3 in context where agents are running around the world on your behalf.

Speaker B: Okay. Yeah. And it's all happening as we speak. We're in sort of really interesting times. So great to have you on on Ash Jake and uh, uh, thoroughly enjoyed this conversation spanning from the time when it was sort of uh, difficult to have a computer ah. In one's home and coding on floppy disk to today, what Agentix forums are really solving today. Uh, thanks for being on.

Speaker A: Yes, thank you so much. Thanks for having.