We Asked Anand Iyer About AI's Biggest Problem

Transcript

what we look at which is there's a lot of money and a lot of capital a lot of talent chasing that pool um and a part of being a VC is is chasing somewhat of a contrarian thesis right or like fish where no one else is fishing right and this is sort of like what got me hooked on AI crypto back in mid2022 when we started making a first low of investments in the space which is I saw the AI through line coming >> how can I think ahead of everybody who understands that demand is is going to keep growing okay great we got that but like what you know how can I formulate a longer or a a second or third order thesis around that idea.

>> Welcome back to the Rowup. We are your hosts. I'm Andy. >> I'm Robbie. And this is your home to learn from the biggest guest, the brightest minds at the tip of innovation. >> Our goal is to help spread the collective knowledge of everyone in our community. >> We are breaking crypto out of its echo chamber and taking it to the real world. >> Sit back, relax, and enjoy a quick word from some of our awesome partners who make this show possible. CryptoUX broken. AI can't trust it. Let Magic Newton set you free. It puts your onchain transactions on autopilot. Does all the work for you 24/7 so you can go do what you want. The vision of Magic Labs, the creators of embedded wallets used and backed by the best. See for

yourself. Simpler crypto UX, verifiable AI agents. Newton.xyz. Introducing Asset Runes, the future of decentralized finance. Stable coins natively on Bitcoin. Hedge against volatility. Diversify your portfolio. All directly from your Bitcoin wallet, secured by ZK Technology. This is asset runes available now exclusively on Starknet.

>> All right, we're we're in the city. It's good to see you again. It's been uh what? It's been 40 years since we last probably spoke even outside of Twitter. >> Yeah, that's right. >> From from the DeFi days all the way now to >> to the AI being the hot topic amongst other innovations in the space. >> Yeah. >> Entire world has changed. >> Um it's good to have you back. >> Yeah, thanks for having me. It's such a treat to be here. Beautiful day in Manhattan. So yeah, good to see you in person too. >> Yeah. So uh when I reached out I I was kind of deep in some AI research work and uh you know our fund has done some AI crypto investing hyperbolic uh mirror some of these types of teams and and you know we we we get it but

like I think there's a lot more to get for our audience as as well as as myself and I really wanted to understand to start your core thesis here as to these two worlds colliding and I'll kind of bring in this compute thing and and demand conversation but first you know what are you doing and what's your core thesis? Yeah, again awesome to be here. You know, this has been developing the thesis, talking about it, working with others who with whom we see eye to eye and kind of living in the future that we all want to be a part of has been sort of like what's been top of mind for me for the last several years.

And you're right, the world has changed a lot in the last four years and probably look drastically different. And hopefully it's not another four years before we can talk again, but it's kind of crazy how quickly things are moving. um at at some point in the last few years since we met um and maybe you know obviously GPT3 was a big inflection moment for for everyone Chad GPT especially um >> I I started to think a little bit more about I guess the what you know very very early on one of my investors had had talked about this thing about living in the future and that's always kind of stuck with me this is Mike Maples from floodgate and that has become something that I've absorbed and internalized which is you know what does that steady state future look like and That's a really hard thing. Those two are kind of

oxymorons in some ways, right? Like steady state is going to keep slipping away from us. The future is going to change keep changing. Um, but the thesis is simple. You know, what does infrastructure look like in a post AGI world? That's that's it. And I know that it's it's not like there's a specific, you know, clock that's ticking towards AGI or anything like that, but uh we're headed there. That's that's the direction we're headed in. that's become that's informed AGI has informed a lot of how I want to think about investment and everything kind of falls under that umbrella for me. So that's been sort of the leading uh I know that's like you know flying at a 30,000 foot altitude but that's been sort of what I've what's been >> informing my investment thesis a lot.

>> Yeah. And Sam Alman actually put out an interesting little blog yesterday speaking about how we're basically at super intelligence now. there are there are some of these AI systems and computer systems which are >> you know smarter than humans and you know he ironically wrote that the world actually isn't as weird as maybe he thought it would be at this current point >> right >> do you resonate with that >> I I definitely do and in fact it's something that I think he tweeted about this as well maybe just coming out of you know 2023 20 you know right right around when the big open AAI shuffle was happen was happening among other things and my timeline might be off here but he said you know we expect this to be you know one of these developments or announcements about a new model and

these capabilities to be very momentous. You know, it's like this is it. Oh my god, the world's going to change and you realize that, you know, it takes it takes takes a long time for this change to happen. Although we're we're growing at, you know, sort of parabolic rates, rates we've never seen before. Uh but it still takes time, right? I mean, sure, Chachi PT is the fifth most visited website in the world as of today. Uh I'm sure that'll change uh maybe by the time you leave here, right? Um but yes, it changes take time, right? And it took so long for us to get used to just seeing Whimos in the city, right? Like it's just like, oh, this is and you know, the amount of driving runs that they've done to get to this point of of autonomy is like is incredible. But now that's

permeating through other cities and and I think that to me that's like a very clear mental model because it's something we've been able to see and physically like watch that progression happen. Um, and now we're here, right? And now it's like, yeah, Tesla robo taxis are are are are happening in Austin. And soon maybe, you know, to me, like I I have a 12-year-old now, and I've always told her that I hope she never has to learn how to drive, right? And I hope her peers never have to learn how to drive. And that was such a big deal for us, you know, like for me growing up, I was like driver's license, everything, right? It won't be the same for her. So, uh, yeah, I I agree. It's going to I think the only thing that's a variable here is that what is that

amount of time that it'll take for us to realize that change? like is it going to be months, days, years like I I don't know what that is, but it it will take time and it'll take you know the the adoption will be gradual but when it happens it'll be pretty beautiful. >> Yeah. Undeniably the acceleration of of this technology is happening and has been happening. I mean you you've been in SF for a number of years. You were lucky to basically be on the ground floor seeing this right before a lot of us kind of re you know had our hands on it and saw it firsthand. Um a and now it just feels like honestly it feels like we're behind but we're just quite still it's just so early in the broader lens right from like from the bird's eye view. Um and you know we you got guys

like Chamath um buying or investing in land out west to uh next to Bill Gates's land to build data centers. Um, and so I've just been continuously refining this idea through conversations with leaders in AI and through our our investment that the demand for compute is just is just going to be continuously increasing um at a scale that is something that we've never seen before for demand for hardware. So hospitals right previously have been the largest uh users of power. I think there something like uh you know a couple hundred gigawatts or kilowatts per hour.

Um they have multiple backup generators. If you look at the grid they are like the biggest source. Data centers now I think are 3xing it. Yeah. >> Absolutely. So >> so yeah so that's happening. So like that I think is a part of the thesis and like the foundation but maybe you can just like lay that out for us that's happening. Why is that happening? >> What is happening there? um you know and and how does that look going forward a bit? Yeah, I mean simply put, it's a it's an arms race at this point, right? And I I'll give you a couple of different Lego blocks, right, as we used to like to call it in DeFi, right? It's very very analogous here, but um one of which is, you know, there was a memo that got leaked from Google maybe a

couple years ago, I want to say, almost exactly two years ago perhaps. It's it's kind of I don't know, it's weird things like NBA finals two years ago. That's the memory that I have, but I could be wrong. But the memo was uh I think the the the top line was something like hey there's no moat when it comes to models, right? Like the a lot of the these a lot of these models are are built differently. They're constructed differently. Like you typically have the very sort of simplistic way to think about machine learning and and models is that you have you have compute, you have data, you have algorithms and you can flip that data compute algorithms, right? Typically. And um when it comes to algorithms is where you can kind of like have have some amount of differentiation when it comes to how you

implement or develop these models. But for the most part they'll end up executing on the same kind of corpus of data because that's what's publicly available. Right? So all these kind of scrape >> you pre-train these models you train these models you fine-tune them. You deliver them and that's kind of where inference tends to happen. But data is is like oil in a lot of ways right? It it ends up being the differentiator. But as you can see, OpenAI ran fast with it. And then of course, we've got, you know, many credible competitors like Anthropic and and Google's models among others that are coming up really quickly.

Mistral of course. And then we have, you know, true open weights models like Mistral's models, like Meta among others as well. Classically, what's happened here is that you have the big incumbents sort of running away with this and they need just large amounts of compute to be able to train these models and continuously train them to get them to a point where it's usable by us by every single day. And it's also somewhat staggering, right? I don't know if you saw this tweet that was going around. I think it went viral on Twitter maybe last week, which was I think it was someone was messaging Chad GPT like it was their personal therapist. And it was like I think you know I think the person ended with like hey thanks a lot lots of love or something like that.

>> I think it was more of like a relationship than a therapist almost. >> Yeah. And I've looked at it so many times and try to process what that means because it's such a personal experience for someone to be able to do that with this thing that suddenly you've gained so much trust in. Um but to be able to do that just coming back to your point and question which is like we needed to train these models to get to this point where it responds very quickly. Well, that's on the infant side, but still like you know to be able to deliver responses that make sense, but then to do them at the clip and pace at which we're expecting them to now >> required a lot of work. And so that's kind of like the amass, you know, like how how much compute that each of these

>> uh companies need to amass like these, you know, hyperscalers need to be able to provide for these models to get trained and then eventually have them delivered so that inference happens against them. that's kind of where the power aggregation's happening and um I mean there's initiatives by Tesla uh there's Stargate by OpenAI and among others purely just to get ahead of this right so um I I'll I'll go off on a slight tangent by saying I think that's that is inevitably happening right like I we talked we talked you know Chad GPT is the fifth most visited website amazing um I don't see that changing um however what's been really top of mind for me is how broad these models are starting to become right like invariably

they they're trying to boil the ocean, right? They want to become the fact that this can be your one-on-one personal therapist or it can answer a complex, you know, math question. You're coming back to the same application to do both of these things, right? And I I have a tough time and I think many others probably have a tough time as well coming to terms with the fact that you can, you know, you can either go broad or you can go deep. You know, trying to do both is going to be invariably hard. So now this is where I think the pivot starts to happen from amassing all this compute which is fantastic if you want to build a truly large language model right these foundation models that are really huge but I think what's coming down the pike and it's already starting to happen are

more more niche more custom smaller language models or maybe not even language models but models that can really cater to a very specific domain that are hyperfocused and trained on this but don't have the same kind of comput requirements because they're not trying to boil the ocean they're not trying to do you know answer questions about, you know, what are weather patterns in in Europe looking like this summer versus like, hey, you know, be my personal therapist. Like, they're not trying to do both. They're only trying to do one or the other. And or just one, right? And that's uh I think that's that's that has a different set of compute requirements. And I also feel like it's more in tune with what we can expect from distributed systems and being able to put the kind of technology we've been building with blockchains to work for a long time. So that's that's a

lot of the thesis that I've been trying to hash out. Like you talked about hyperbolic, it's a great example, right? All of hyperbolic infrastructure is based on distributed machines and data centers that you know are are not necessarily homogeneous, right? So that's a power that's that's a strong suit of of what distributed systems have to offer. So um yeah, I think that's you know two thoughts there, but I think you know I think computer requirements are going to continue to grow from the hypers from the hyperscalers and the large language models. I also think we're going to have a net new crop of models emerge starting imminently that are maybe either challenge neural networks or our or small language models um or like very very niche in nature.

when you speak about the intimate relationship possible with some of these LLMs, I think that that kind of brings in some of the uh elements of where uh blockchain and some of the privacy and data privacy and also kind of guarantees one can have comes in also with uh you know plugging in a prompt and getting a result back kind of verifying like or knowing at least uh succinctly what the the training data was on that given model. Um so like decentralized compute, decentralized inference, uh specialized data, these are kind of the sectors where where we've seen these two worlds collide.

>> But let's put that on pause for a sec and come to the second and third order effects of uh this kind of binary uh demand for compute. So, we've got >> just large demand happening for just anybody who's going to try to build, you know, think about like a a local LM even for yourself or for your company or for, you know, for a large uh competitor to one of the big ones. Now, then you've got, you know, very specialized compute needs for uh you know, a a very kind of like vertical specific use case. So, these two things happen simultaneously.

the cost of power um is I I assume energy has been going up through inflation and just through this overall demand. Um I is is there is there a second and third order consequence of this binary outcome that you're explaining or like yeah >> basically what I'm getting at is like okay how can I think ahead of everybody who understands that demand is is going to keep growing okay great we got that but like what you know how can I formulate a longer or a second or third order thesis around that idea I think that's what you've been doing basically at least in in in in some capacity.

Yeah. >> Yeah. No you're I think that's the right way to think about it. I think there's the um there's the more immediate effects of this and you're right like power consumption has has gone through the roof. We were talking, we were riffing on this before we we could start rolling, but you know, I think I think the forecast, if I remember my data correctly, is that like we'll we'll basically between 2023 and and like by next year, we'll have double the amount of power consumption that we'll have the amount of power we consumed in 2023 will be half of what we'll have consumed in 2026, right? So, and I believe like an overwhelming majority of that maybe more than 50% of that will be consumed by AI, AI needs, right? which is kind of ridiculous to think about. So that's um that's sort of imminent and here and now

and it's happening. Um, you may have seen this news article flash by I think it was last week or the week before where there was this, you know, Meta was talking about how they want to colllocate some nuclear reactors next to their data centers, right? So to be able to harness some amount of clean energy and so that's another interesting uh, you know, maybe a second order effect, right? like when we talk about I think what you're alluding to which is which is really uh important and poignant here is that we we talk about like hey the GDP growth that's going to come with AI is like so is going to be like insane by the time we get to 2030 and this is the reason why right like it's it's not about like you know us paying X you know $20 a month for Chachi PT I think that's

one thing but underlying that is just so much work that's going into this right um like you said the the the acquisition of data and like training of models and energy requirements all the way down to the chip level right like to be able to create fabs and chips in you know what are the what are the effects of tariffs right like that's another uh effect that we should consider should we be thinking about you know collocation of some of these industries in America or in in uh in South America and Latin America right so that's uh all these are like really interesting effects to think about and then uh so yeah that's that's like the through line that's like super clear but when it comes to aggregation and and I mean you know when it comes to the truly

large language models right and then yeah the other piece which is um now to apply the venture lens right like you know very um you know somewhat selfishly and egocentric >> how to make money on this idea right >> yeah or at least how to think about it more strategically from a calculated risk perspective yeah >> yeah and I think it's truly like you know to speak you know more more cryptonative here right like left curve right curve right like I think it's um I think we're going to have I don't think I I Oh, we're going to run away with these large language models, hyperscalers and that that just need a tremendous amount of venture capital to be able to be successful. And the game they're playing is different, right?

Like it's we think they think that this is going to be, you know, a trillion dollar outcome, right? It's we're not playing for a deck of corn status anymore. Like it's a we're or a couple orders of magnitude away from that level as like the benchmark now, right? Um and that's fantastic. like the kind of outcomes we'll see you know through that kind of venture investing is going to be different but also requires a large amount of capital right um and I I I was looking into this a little bit which is I think in the last since 2022 >> through till maybe March of 2025 something like $300 billion has been invested into AI startups completely that are not in the AI crypto domain right so purely like in quotequote centralized AI so something like that >> um and loosely speaking today like if

you had to do some amount of like back of the napkin math based on public available data those companies altogether today are worth about 1.5 trillion right so that's sort of like the state of state of the market today now you know talking about coming back to the venture side of things like our what we look at which is there's a lot of money and a lot of capital a lot of talent chasing that pool um and part of being a VC is is chasing somewhat of a contrarian thesis right or like fish where no one else is fishing right and this is sort of like what got me hooked on AI crypto back in mid 2022 when we started making a first of investments in the case which is I saw the AI through line coming. I thought this intersection is going to be really interesting. Um

and to me the story arc was we started off by showcasing that crypto is excellent for store of value. This is when Bitcoin first got created. We showcased with Ethereum and Salana and and layer 2 solutions that we can build a decentralized finance system on this, right? Which is how you and I got connected, right? So that's our second arc. And now we're in the third arc, right? Which is AI, right? And that's what's going to drive utility and and really all kinds of innovation with blockchain technology period. So that's what I'm focused on and I think it's super early. Um and I've got a chart that I'll share with you. I kind of compare sort of logarithmic growth of market cap with Bitcoin, Ethereum, Salana from the time they got spawned till today. And AI is on like this, you know, AI tokens are on this upward curve, you

know, compared to Bitcoin, Ethereum, and Salana. So I think we're off to a really strong start. >> Um but it's still very early days. And so that's where I'm focused. And so like so that's that's one I guess the second point in this uh you know decentralized AI thesis is that um there there's a lot of work still to do here. It doesn't look like large language models because um that's I think I feel like the hyperscalers and the the large LLMs are running you know foundational companies are running away with it but we have proven what we've proven so far even like we have decentralized training companies like Jensen that have are able to do things like reinforcement learning using you know heterogeneous hardware across the globe or what you know noosees and prime intellect have proven by uh by training a model across uh

geographies right which is all we thought was impossible to do a couple years ago. So, we've kind of validated that research, but still for LLMs, right? So, for me, the next arc is yes, very niche, very custom, small LLM, uh, you know, maybe we call them SLMs, right? Um, that don't require the same kind of compute or I I'll I'll take it one step further, which is again there's a, you know, maybe a corpus of the research community is talking about this already is all these models, as you know, are are basically transformer-based, right? So, what comes after transformers, right? Like what is that post transformer era looking like? And wouldn't it be beautiful if that was born with like blockchain first principles, right? Like that's something that I've been thinking a lot more about. So that's where my VC mind's been going a lot and where I've been spending

a lot of time. >> Okay. So what part of what I've been struggling with is trying to understand um exactly >> like the need for certain elements that uh blockchain or this idea of distributed systems brings to the AI stack. I know that we talked about, you know, I mentioned some of the data privacy things with, you know, there's companies like Nillian doing like data privacy for for AI work. Um, then then I think about, okay, so when I type in a query, >> when I type in a good, >> that's good. >> Thank you. >> When I type in a query, I'm like, okay, um, I I don't want this information to be leaked. I also don't necessarily want that to go to um, you know, these big corporations. Okay. So, that's that's that part which I think that you I'm sure you have some thoughts on. And then

then when I type in something, I I want to know that the answer that I'm getting is not wrong or that there's no like, you know, large hallucination happening. So, how do I know that? So, there's probably something applicable there. >> And then I also want to be able to have uh I I I want to be able to probably contribute to training my own model. Yeah. >> Or like have >> 100% >> have a system where I can like rely on other smart people to help me train a model and like reward them for that. So, there's like three use cases that I'm thinking about. Maybe your book of these particularly compelling what am I missing in this intersection here?

>> Yeah. The second question first which is you're you're spot on, right? Like this is um this is this is that steady state. we will get there and I think you know Ilia from Near has been talking about this for a very long time that you know what is you know we we all have these LLMs but whether these models running on the edge or they're running on our devices you know kind of like maybe not too dissimilar from what Apple's been talking about for some time this is this is happening this is happening sooner than we think you and I may be able to talk about the fact that like hey we want to have our personalized models but for the average user as you know like they don't know what that is right they have some interface that's hyper customized for And that could be their

own model. That could be a that could be an agent that's talking to some uh very customized fine-tuned model. It could be a swarm of agents. It could be any one of those things. But yes, you know, it it does it does build on these core blockchain primitives and principles, which is, you know, what we've really come to, you know, adapt sorry, adopt in a lot of ways, which is, >> yeah, privacy, privacy preservation, um, sovereignty, um, and a sense of ownership, right, when it comes to these things. And so, I'll zoom out by saying to like let's just start with this as a thesis and something we want to prove to.

What if AGI were something we could stake against, right? What if we could have a say in AGI? What if we could own AGI, >> right? >> I I don't like for all of us who were smart and involved in the space today, >> none of us can actually concretely say that we're moving in that direction. >> No, >> this concept of AGI is something that the larger companies are running away with today, >> right? >> By long shot. Yeah. >> Yeah. By by long shot. So if we want to work towards this concept of ownership of AI then that needs to start now and it starts with all the tenets that you just mentioned right so privacy preservation um and sovereignty like these are like just like core simple tenets understood >> and so that all that being said um we're

working towards it I think it's you know this is the the classic scurve of like how technology evolves right so to me it looks like >> we started off by showcasing that blockchains can like can can survive that we can do distributed systems with with token incentives at scale >> and then we said okay maybe it doesn't scale so we need to scale it so we had all these scaling solutions that came up and then we started to build on top of that right and specifically in the AI domain we start to look at things like decentralized compute with with companies and projects like Akos um we looked at you know there's companies like VA that can do data collection or aggregation at scale companies like Sahara that can do data labeling um and then uh and then we have companies like hyperbolic that can do inference using

you know distributed network of of machines and so but that's the arc that's where we are in the S-curve and I think the fourth block is where we start to inch towards a sense of ownership and and maybe it it kind of collides with the sense of um AGI at that point >> so so you know Ilia would call it userowned AI >> right >> communityowned AI but effectively the and it almost sounds like >> not altruistic but it sounds like um you're fighting for a good cause in a sense right as part of the thesis, right? It's not just like it's not like like the the thesis is centered around we don't want the future of AGI to be owned by you know a set of technocrats if you will um and so you know you put

these different pieces together hyperbol like Vanna etc and you you can create a system potentially that is that can rival that of something like a like a anthropic or like a uh you know like a Google right >> right >> so so like so there's kind of like a feelood good cause here going or not a feel good but it's actually more much more dire than that >> if you think about it in a dystopian manner. >> Sure. >> Yeah. Introducing Summerfi, one of DeFi's oldest and most battle tested platforms. Summerfy automates your exposure to DeFi's highest quality yield so you can sit back and enjoy this summer. Deposit stable coins at USDC or USDT and let Summerfi route your funds into optimized low-risk lending

strategies. With over a 100 million in instant liquidity and support for over 34 protocols across Ethereum, Arbitum, and more, you always stay in control. Choose from a wide range of curated strategies reviewed by independent risk experts, all in just a few clicks. So why work hard for your yield when you can be lazy? Start earning now at summer.fi, the home of the lazy summer protocol. Innovating finance, global, institutional, and onchain. Deep liquidity to drive capital efficiency. Now maximize yields on institutionalgrade liquid staking with MH.

Unlock trillion dollars of capital via F Bitcoin. All of this secured by Mantle's treasury of billions in stables and ETH capital pools. This entire ecosystem is coordinated by a vast network of seasoned VCs that are aiming to create the next generation of banking and financial technology. This is mantle. >> It's it is and that's the scary part, right? Like I'm not I'm not like, you know, a doomer trying to Yeah. But >> but it's a valid point, right? Like I mean let's going back to the example of this you know people and by the way this is happening pervasively. I don't know if you noticed this there's a trend here. I couldn't tell you this you know exactly quantitatively but qualitatively

I've heard from friends uh whose children who are like teenagers starting to use chatbt as a true you know uh relationship coach right like very intimate details like what they'll do is this is I maybe there's already a story out on this and I've missed it but they're taking screenshots of messages they get from their significant others right girlfriends or boyfriends and they're just literally pasting those into chat GPT saying like what did she or he mean by this like what does that mean like what What does it mean when he says X Y and Z? Right? Like isn't that crazy? >> And um >> not that these systems are that sophisticated, right? Like they're not trained on like you know Gen Alpha, you know, slang. And so then you know like this is tough to begin with, but >> um now like okay, so that's that's

happening, but imagine for a second like you pointed out hallucinations are one. So it's like you know it's a it's a uh in like when the stakes are low, right? Like it could be a benign side effect, right? But when the stakes are high and you have uh you know something like let's say there's some amount of evil that's kind of snuck into this. I don't know if you saw this happen on Twitter with Grock maybe a few weeks ago where it was just talking about the apartheid like you know constantly right and like that it was it was a malicious intent right so when you have that when you're in that quadrant of like you know hallucination gone arai and there's some like mal intent behind it now suddenly imagine like you know I don't know teenagers who are using this for

relationship advice are getting influenced >> by a large corporation >> right so I think that's the control right like and this This is what we were I guess in a lot of ways at least me as a parent like I think about a lot of when it comes to social media and that influence right so I mean there's no like in my in our household there's no Tik Tok there's no Instagram there's you know so >> reminds you of my dad growing up >> yeah and so um because it you know it's it's scary to think about I mean my daughter's old enough she can make her make her own choices and decisions but for now we're like you know let's let's try to do our best by you know as as good as good parents um but all this to say that like you know but she has

access to Chachi PT right >> uh and Claude And so what's uh what's to say that you know that those might not be an influencing factor for her at some point. >> And so yeah, it's um >> this is this is like beautiful science fiction. This there's there's a book about like the diamond age, you know, talks about this like the the illustrated primer which is this device that this girl sort of grows up with. Uh you know, like it's it's so classic. This was written back in the day and I think we're really playing it out now. So yeah, the worst case dystopian and the best case we have a sense of control and ownership over something that's so personal to us.

>> Yeah. I mean it feels more abstract as if you you can imagine a super intelligent some some sort of being. It's almost feels godlike in a sense and then you take that and you you assess who has control over that which could effectively have control over the world's communications in some capacity social networks information etc. >> And and it becomes clear that the need for uh again some sort of decentralized ownership of this um of models of training etc. So that like we we got the thesis down but like what are the challenges that these companies are facing um cuz it can't be an easy battle for you know Jensen and Hyperbolic and Vanna and Ritual and and u you know

insert you know portfolio company here and going and looking at the you know the the battleship across the way that that is anthropic or that is you know a cursor whatever it is. um what are the challenges that some of these like AI crypto companies are facing right now that you're seeing um yeah >> in their in their kind of operation? >> Yeah, it's that's a it's a great question. Um turns out that if you by the way, you know, the the research that went into OpenAI was like a decade long, right? So they had such an interesting head start and um look I was I was walking over here with a friend you know a mutual friend you met him as well but I was telling him that I I you know everyone in the valley like invariably

you've met Sam Alman and like you know what a interesting thoughtful forward-looking uh guy and I remember meeting him in 2007 and there's something because he was he was actually building a company back then and he said that my goal is to build something that everyone in the world would eventually want to use right like and that's that's been with him for the last like almost 20 years now, right? Like that he's been thinking about company building, right? So that's that's his that's his reason for being um and so he's had this head start. He's been thinking about this for 10 years um you know through open AI now. Uh so we're playing a little bit of catch-up when it comes to uh not only building large language models like they've been doing it but doing it in a decentralized fashion >> number one um or in a distributed

fashion number one but also on a on like heterogeneous hardware right so like the reason as we were talking about like you can amass all this comput you can go raise you know tens of billions maybe hundreds of billions of dollars right to go be able to build these massive clusters and do training against it um that's one way to do it but we're trying to do it the hard way right this research really hard, right? And so and then to to try to prove it, to make it work, that that takes a lot of really smart people, the right capital, the right time, and all the sort of like ingredients to come together to make the perfect sort of formula. So I'd say that um it's happening, you know, and a couple years ago, everyone was challenging that, hey, decentralized training cannot happen.

>> We've proven that actually it can happen now. Um we, you know, and that is that is actually the crux of it. That's actually like the was was one of the bigger blockers when it came to thinking about building models that we feel a sense of ownership around like that has to that's like the first lead into almost everything else. So, I think we've gotten there, but to answer your question, I'd say what we don't have is like the luxury of time, you know, because we've got resources, right? >> Yeah. And and resources and it's happening, you know, and this is why, you know, crypto VC as I was mentioning, right? Like $200 billion, $300 billion invested into centralized AI, roughly a billion into decentralized AI in the exact same time frame, right? Like that's a couple of orders of magnitude less. Um, and we're still expected to

catch up here somehow, right? Like that's interesting. So, um, it's it's a tough one and I think so we're a little bit behind and so >> I guess my call to action would be, >> you know, more more talent, more resources. The good news is that like, you know, you mentioned Jensen and and Hyperbolic as two examples, right? The founders of both those teams are exceptionally smart people like you know they're PhDs and ML and they've been thinking about this problem for a long time. I mean, yeah. So again, not to talk the book too much, but like these are those are the kind of people we need more of and then with the right capital infusion, right time, I I think there's a lot we can do, but yeah, I I think we are coming from behind a little bit and but good news is that we have this

platform on top of we can build now and that's got that's kind of like what we needed to get to and like as I was mentioning in the S-curve, we're like there now. So now we're on this launch pad to go off to crazy things. So we'll see how that pans out. I think uh similarly to how stable coins have had such a positive uh perception change and impact on crypto as a whole as of recently uh there will need to be some sort of sentiment and realization more broadly at a macro scale of the >> I guess negative consequences of let's call it traditional AI or centralized AI but that's a massive education gap from even people using the tools today to then realizing the downstream effects of the tools to then coming back to the idea of oh there there should there's a

better way to do that. So I think it's also just a big time thing, right? >> Yeah. >> But as you say, time's kind of against this in in in a similar right. >> Yeah. That's a great analog by the way, right? Like I think there have been folks who've been talking about stable coins for like almost six or seven years, right? And like uh it it took this amount of effort and um scale and also you we had some headwinds over the last several years and that's changed with you know uh the administration and um you know more positivity when it comes to regulation among other things.

So all that's great like I think the timing feels good for it. But that being said, hopefully yeah, we we've we're coming from behind a little bit, but I don't think it's, you know, it's not too late because I I still do believe that um and just to zoom back out again to go back to the venture play, right? It does, you know, to have a winner in centralized AI is going to take a lot now, right? Like cuz the bar has gone up so much, right? I think if you're u if you're trying to build a company that's not going to be a decacorn at the very least, right? like, you know, venture is going to look a little bit different for you. And uh so I hope that doesn't styy or stifle the people who want to build there. And by the way, that's another

whole thing I'd love to talk to you about because I think that's there's a whole other play for crypto there that's that's unrealized yet, but I think it'll start to become more prominent very soon. But yeah, that's the that's the game we're playing, right? Like and this is the beauty of it is that there's those who invest in OpenAI and anthropic are going to win beautifully, right? That's going to be great. But uh that hopefully that doesn't mean that we sty innovation when it comes to everyone who's not building an LLM or a foundation model because we need more innovation across the board. >> Well uh you've you've you've uh piqued my curiosity. >> Yeah. >> Um so that there's a there's a sector certain developments that you're seeing in SF in the more traditional AI space that you think could be >> Yeah.

>> could be benefited from you know having more kind of crypto AI ideals and technology. >> Yeah. Um it's simple. I'll put some like bare stats in front of you, right? So, um, and first I'll maybe start with some personal activives, right? Like there's there's some great sites that are now tracking how there are purely bootstrap companies that are basically putting LLMs to work to create products that maybe couldn't exist before, right? Um, uh, and these are everything from, uh, you know, like there's definitely a long tale of like coding related apps. So, you mentioned cursive, but there's a whole slew of others. um to um like uh you know uh fitness personal health

related apps that are like LLM rappers uh to like very niche in insurance use cases and these are all like purely bootstrapped companies that haven't raised any amount of venture financing but they're actually doing quite well they're actually trending well in terms of millions in ARR and um now the question is like do we go raise venture capital and if so you know how much we raise and like what's the comp here like how do we Should we think about this right like what do you displace and what's the outcome and among other things right so but these companies should exist because they actually have users and they you know this innovation should happen but the traditional venture model might not work for them right >> they want Yeah. Yeah. And for investors too.

>> Yeah. >> I think what what you're getting at is like they should launch a token. >> Yeah. They should they should >> like there's a better path to liquidity. >> Exactly. Yeah. And that's tokenization. Right. I saw a massive uh investor speaking to the the GP of uh Vault Capital on a podcast. >> I don't remember his his name, but his his >> Bill Bill Gurley. >> Bill Gurly. >> Yeah, Bill Gly was talking on on on Colossus to Patrick Ronessy about this. Um and yeah, >> and his quote was that the crypto is now one of the only viable paths to legitimate liquidity in venture now. Yeah, basically some some quote like that.

>> Yep, that's right. So what what you're getting at is that and I had Pranov um from VANC, he's running their liquid side here, had him on last week. He's talking a lot about uh the onchain IPO era. >> Yeah. >> Or he calls it tokenized equities. I think that's a bit of a misnomer because people think like, oh, we're going to bring Tesla on chain, but no, it's we're going to take your equity that you've built. >> Maybe you're not like a $500 million a year company. Maybe you're a $20 million a year or a 10 million and you're just going to go and you're going to raise with a token comp and you're going to launch a token route rather than having to go that route.

>> That's right. Yeah. Yeah. And we should be Yeah. We should be careful about the word raise, right? Because what like you know uh I guess Commissioner Pierce Hester Pierce has been talking about which is like we need better clarity around this because when you call it a raise and suddenly you're you know you're >> bordering on some securities issues that you don't want to breach yet. But uh it's true right like what is what could be a viable path here like again the options are one you stay indie you're like you know you don't need to raise any venture you don't need to have any liquidity or a lifestyle business >> you make money and that's fantastic that's great too like that's the beauty of innovation America like we should be able to do that >> and the other end you know become you

become a massive success right like you have do you like world domination is the goal right like I want everyone in the world to be able to use my app and then there's some middle ground here maybe closer to the right than it is to the left which is um in terms of like the venture scale versus like being truly indie and uh yeah maybe tokenization is the angle but what is crypto really good at crypto is a beautiful capital coordination tool like that's what it's historically been amazing at so let's put that to work and let's make sure the right guardrails to be able to do that um I'll add a second thing right which is um which I think was and I think Bill Gurley didn't get a chance to allude to in his talk with Patricky but um there is actual true retail demand

for good tokens in crypto >> like a lot. >> Yeah. >> Like like it would be great. You know what I mean? >> Exactly. And so uh there's there is a little bit of fatigue if you look at I mean maybe the last several years is not a good indication because we didn't have a lot of tokens getting listed by major exchanges but there is some amount of fatigue there uh when it comes to >> another layer 2 or another scale or another data availability product or like I mean know it's too it's too unclear right but if you just look at the performance of AI tokens right over the last year like from May of 2024 to May of 2025 those tokens have basically doubled in in market cap between May of 2024 and May 2025.

>> Um during that same time, Bitcoin was up 46%. Everything else was literally down or negative, right? Including memecoins, right? Um and so this is there's clear appetite. By the way, I'm not saying that only nuance in in in that picture by the way, but >> Oh, a heavy nuance, right? It it was heavily >> there's always outliers. >> Yeah. And to both sides, you know. >> Yeah. And I'll also fully admit that those were there was two primary drivers in that category. One of them was virtuals, the other one was bit tensor, right? These are the biggest drivers of that uh of that drive in value. But for but that's also somewhat power lawesque, right? Like you can see that playing out.

>> It's dredicated to a larger idea, right? >> Yeah. There's there's appetite for it. But >> if that is just, you know, let's take crypto AI tokens aside for a second. Uh there was a there was a study that was done by Charles Schwab. Uh it was an ETF survey and I think it was done late last year. Um, I think it was something like 62% of millennials that were pulled want to have access to crypto ETFs and something like 47% of them want to have access to be able to invest in AI, right? Like that's just invest in AI. >> Yeah. Like Yeah. Like we've done like four or five checks to these AI crypto companies, but I have I have zero exposure to I I I got the X AI deal >> and it was like 5 billion y >> and I was like, dude, like what is my

25k check in doing this $5 billion valuation? like, you know what I mean? >> Yeah. >> So, people want access to those deals and right now they're they're they're out of reach for even somebody who's, you know, semi successful and has capital like like us, not at large scale. Um, you know, that's basically only available to the best of the best out west. >> Yeah, that's right. Yeah. And even those are like even those that might be available like SPVS of SPVS or something like that, right? Which is like, you know, by the time your fees are get, you know, get get taken out, it's just like, you know, what are you really getting a hold of, right? So yeah, there is there's this there there's there is a lot of leverage that comes with having a large platform to be able to deploy large amounts of

capital and really stand behind that with conviction, right? That's one thing. But yes, the >> yeah the average investor wants to be able to invest in AI. What are they going to do? They're going to invest in Nvidia, >> you know, maybe Coreiv now. Cor isn't doing well. uh Google >> and and also the average founder wants to >> be able to have a liquidity event. >> Yes. You know what I mean? >> That's right. Yeah. Yeah. >> Like you want to have something and the likelihood of an IPO is I mean >> it's been so hard. It's going to just keep being harder. I mean I would imagine or just being the same level of difficulty.

>> So you have two kind of forces that are going in the same direction. >> 100%. You have demand for these kinds of tokens for the forget tokens. We have demand for AI right >> and businesses that are generating revenue. >> Yes. Yeah. So I I believe this is going to catalyze a whole slew of different tokens coming to market. >> Very interesting. Prof said the same thing. I was like so reading it old season. He was like, well, >> there's going to be an old season, but it's not going to be the way you think. >> It's going to be these these onchain launches for these profitable businesses that come and launch on chain. >> That's right. Yeah. And we're we're yet to today this looks like again we're playing with this a little bit, right?

looks like agents that are launching on virtuals or projects that are using the believe model, right? >> Internet capital markets idea, right? >> Exactly. Yeah. And uh monetizing through but I think we're showcasing that this is possible today and invariably this is going to become an instrument that's just become second nature. And when that when that gets into effect like you know this the marketplace dynamics are going to be absolutely beautiful. And by the way uh I Bill Gurley is like a revered marketplace investor like if you go back and look at his articles um uh I I think his blog you know both sides of the table is like inc. Sorry that's that's not his blog. I'm sorry I got it wrong but his his posts are just like absolutely phenomenal when it comes to

marketbased dynamics. So he's thought this through. He's he's uh I'm very lucky to have him as an LP in our fund but he's a very very smart guy. Yeah. Yeah, when I read that I was like, okay, I think we've been doing the right thing with with our time for the last couple years because these uh you know, doing the whole 8 to 10 year liquidity cycle right now for some of these things. And I mean, also, crypto's just had this asymmetric return profile that is probably going to like slow down a bit in terms of just like the pure vapor that has done well. But I think that's going to be as a result of basically what we're talking about here.

>> So, so, okay. So you're in the trenches then in out out west >> finding quality AI founders and persuading them or at least conversing with them >> about potentially doing this type of model rather than trying to do the traditional model. >> Yeah. Um and honestly like I I I don't think you know as you know like you have to be you have to have your foot on the door. You have to be able to have these conversations and >> I mean credit to all the smart people who are building companies in San Francisco. I I don't feel like a pariah there. It's not like, oh, the crypto person's enter the room or anything like that. Like, it's there's a lot of like open-minded, you know.

>> Yeah, I bet. I mean, it's pretty open there, isn't it? >> It's very very much so. And I really appreciate that. I'm not saying that, you know, there aren't these bubbles, right? But there, you know, to me, I feel very accepted and and so that's the good news. And I I've I've lived there forever, right? So, to me, this is like this is first nature and I appreciate that. It is an echo chamber. I'll be honest with you, like it is its own bubble for sure. >> Like crypto, a crypto echo chamber is in New York, right? the AI economies in San Francisco and I like having this contrarian hat where I'm like I'm sitting there thinking about crypto as it relates to AI and over here we're thinking about you know it's like crypto is sort of first nature. So, um, but yeah, I I I think I'm I'm not at that

point, I guess, just to respond to your comment. Like, I don't think we're there yet to persuade anyone. And I think >> the minute that some of this starts to like truly ignite in some folks that >> I I don't, you know, I don't need to subscribe to this traditional mindset because I don't think we've I don't think we've got the recycling yet, right? Where like venture is still flowing in to AI startups like that's not not happening. But the minute we realize that it is the power law is so drastic here that you know you have to hit a I don't know trillion dollar outcome for some of these funds to get returned or like tens of billions you know it's it's going to it's going to change how they think about investing a little bit. So um >> I mean I think the founders would would

be would uh would be doing themselves a service assuming the crypto market dynamics remain strong and intact structurally. Yeah. and the lack of, you know, let's just be a little bit pessimistic, the lack of quality uh businesses being that have tokens that are, you know, >> tied to a direct business that has long forecasted revenue. >> Yeah. >> You know, it's like it's kind of a blue ocean there where whereas the the general VC seems to be like it's not crowded by any means, but it's a bit more of a competitive landscape. Yeah.

in terms of raising raising or you know getting capital and then launching. >> Yeah, >> it's definitely crowded. Like I mean it is like >> it's crowded as >> Yeah, it really is. Yeah. I mean there's >> Yeah, it's good. It's good and bad, right? So yeah, that's Yeah, that's the state of affairs today, but I hope we see more of those kinds of clever instruments being in, you know, really used and implemented and deployed uh the way we expect them to. >> All right, we're going to end this on some light-hearted stuff. We're going to react to a couple of your tweets. All right. >> All right. Oh All right. >> All right. Here we go. Um, uh, here we go. You got some good ones coming up. We got a lot of retweets here for you. I'm going back. I I'm often asked why AI agents need to be

tokenized. It's because crypto can price what traditional markets cannot. Tokenization lets us trade, own, and incentivize AI and agents directly by creating liquid markets. This has and always will be crypto superpower. Do you get in invited to the no coiner AI parties in SF from chain Yoda? Haha. AI party. All right. So, crypto's ability to do capital coordination and formation at scale >> is an incentivization mech kind of what we're just talking about here. >> Exactly. Yeah. >> Yeah.

>> Yeah. >> It is it's a little um I I think there is this disparity between what you expect something to be worth like in our minds like how do you how do you know what some like you know it's like it's always I've always had this issue too. I don't know how you thought about NFTs, but I've had this issue like understanding how how to price an NFT, right? And to to some extent some agents that aren't quite at that point of any delivering any kind of value, right? Like there's there's some hype or some mania behind it to be able to value it at something. >> And um I think that has been crypto's trend, right, in some ways. But that's what we've done. We provided liquidity to these assets. That's kind of been weird or hard to assess in some ways, right? So I I I do think it's a

superpower. I think we lean into it. Um that doesn't mean we don't get this these outliers of of uh of uh you know stuff sometimes, but that's how we learn >> stuff. Yeah. Yeah. That's kind of like the what the last Asian memecoin thing was about. >> Yeah, that's right. >> But the core idea remains. >> Yes. >> Probably believe in it. >> All right. Here we go. We got another another good one here. Um >> it's like having your diary being read out loud. I'm like, "Oh, wow." Except it's public election. >> We are entering a regime where the product is the raise and the market is the funer. We're living in a world of interoperable capital markets. MVP plus feebased traction is better than a pitch deck. Agent fees fund runway. OTC liquidity replaces series B. Narrative

times community time mechanism equals escape velocity. What this means the next $1 billion project might be vibecoded by a single dev or two launched on virtuals and sustained by fees and a lot more about the kind of capital formation moving onchain. >> Yeah, >> walk us through these interoperable capital markets. What's the what's the thought process here? >> Yeah. Um, I I do have this call call to action that there will be a billion dollar market cap project that's been coded by maybe one person, >> right? like I I it's in inevitably going to happen in my opinion and it'll happen in this cycle and maybe there'll be more than one >> um and there'll be a state where this is

created maybe spawned by another agent right that's developed something and so uh like that's that's the living in the future piece of like you know trillions of agents all having their own economies and pricing one another >> um but yeah it's uh that's one right so we talked about this a little bit earlier which is you know do we how should we challenge our consensus is thinking about venture style fundraising and the token tax model which I think is is really innovative and it's it's working in a lot of ways for products like virtuals um is is like a really fascinating concept right like you're truly getting it's one one a one source of revenue by the way but the token tax can fuel innovation can can fuel your development which is kind of cool it's

your it's it's an you know it's an augmentation to your ARR not going to replace it >> but it's a great way to get boots strapped and as long as you stay honest and the project keeps growing and thriving your token drives in value and invariably you know the taxes get acred back to you and so um yeah I I'm excited I'm excited for the world in which we have these projects that are going to deliver immense value that get priced accordingly but there's a very skinny team behind it and they don't need to raise a lot of venture capital. >> Yeah and I really appreciate your time man. It's been a great great conversation, lots of insights and just fun.

>> Yeah. Thanks so much for having me on. It's fun great to finally meet you in person and do this together. >> Yeah, man. Cheers. Thank you.