AI ROLLUP #13: Crypto AI vs Silicon Valley | Live From ETH Denver
55:28 Share

Welcome, Bankless Nation, to the AI Rollup where we stay up to speed with the emerging trends developments in the AI crypto space. I'm David Hoffman here on the ground floor of ETH Denver. We did this episode, this AI Rollup with Ejaz in person, but we also did it with Tom Shaughnessy. So this is going to be a little bit different than our usual weekly recap episode. It's more of a discussion.

around the current AI crypto sector. Joined by Tom, a new face from Delphi Digital. Before we get into the episode, we're going to talk to our friends and sponsors over at Reserve Protocol, who just introduced decentralized token folios, DTFs, which are like ETFs, but for crypto. There are indices across all different sectors that you can find inside of the crypto space, real world assets, DeFi, even AI agents, virtuals. Check them out. There is a link in the show notes. Tom,

Tom Shaughnessy from Delvi Digital. They've been investing in AI for a very long time, since well before the AI agent meta really took root in last November. They have been investing in the crypto AI intersection for over two years now. So they've actually already seen a number of waves of investing cycles. And so as young as this space is, Tom is one of the few veterans out there. So Ejaz and I had a fun time just chatting with Tom.

just about everything that he's seen so far in the AI space, what he thinks about it, how he thinks about investing in this space, and what he thinks is on the horizon. So let's go ahead and get into that conversation with Tom and Ijaz. But first, a message from some of these fantastic sponsors that make this show possible. In the wild west of DeFi, stability and innovation are everything.

which is why you should check out Frax Finance, the protocol revolutionizing stablecoins, DeFi, and Rolex. The core of Frax Finance is FraxUSD, which is backed by BlackRock's institutional Biddle Fund. Frax designed FraxUSD for best-in-class yields across DeFi, T-bills, and carry trade returns all in one. Just head to Frax.com, then stake it to earn some of the best yields in DeFi.

Want even more? Bridge your Frax USD over to the Fraxdal Layer 2 for the same yield plus Fraxdal points and explore Fraxdal's diverse Layer 2 ecosystem with protocols like Curve, Convex, and more, all rewarding early adopters. Frax isn't just a protocol, it's a digital nation powered by the FXS token and governed by its global community.

Acquire FXS through Frax.com or your go-to DEX, stake it and help shape Frax Nation's future. Ready to join the forefront of DeFi? Visit Frax.com now to start earning with FraxUSD and staked FraxUSD. And for Bankless listeners, you can use Frax.com slash r slash Bankless when bridging to Fraxel for exclusive Fraxel perks and boosted rewards. Imagine a world where your day-to-day banking runs on a blockchain. That's exactly what Mantle is building.

powered by a $4 billion treasury and poised to become the largest sustainable on-chain financial hub. As part of their 2025 expansion, Mantle is introducing three new core innovation pillars that bridge traditional finance with decentralized technology. First is their Enhanced Index Fund, aiming for $1 billion in AUM by Q1. It provides optimized exposure to Bitcoin, ETH, Solana, and USC, complete with built-in yield opportunities.

Next, Mantle Banking promises to revolutionize global value transfer through seamless blockchain-powered banking services, bridging crypto into your daily life. Finally, Mantle X blends AI with DeFi to deliver an intelligent, user-friendly experience for everyone. And the best part is that this is all in addition to their already launched products like Mantle Network, ME, and FBTC. Ready to step into the future of finance? Follow Mantle on X at Mantle underscore official and join the on-chain revolution today.

The Arbitrum Portal is your one-stop hub to entering the Ethereum ecosystem. With over 800 apps, Arbitrum offers something for everyone. Dive into the epicenter of DeFi, where advanced trading, lending, and staking platforms are redefining how we interact with money. Explore Arbitrum's rapidly growing gaming hub, from immersed role-playing games, fast-paced fantasy MMOs, to casual luck battle mobile games.

Move assets effortlessly between chains and access the ecosystem with ease via Arbitrum's expansive network of bridges and onrifts. Step into Arbitrum's flourishing NFT and creator space where artists, collectors, and social converge and support your favorite streamers all on-chain. Find new and trending apps and learn how to earn rewards across the Arbitrum ecosystem with limited time campaigns from your favorite projects. Empower your future with Arbitrum. Visit portal.arbitrum.io to find out what's next on your Web3 journey.

Thanks to Morpho's immutable code, you can rely on the gold standard of decentralization, giving you and your projects full ownership and control.

And if you're a user wanting to earn or borrow on your terms, Morpho's newly refreshed app at app.morpho.org makes it simple to tailor your own risk reward profile. Choose from different vaults and markets curated by trusted third parties. Plus, Coinbase users get access to crypto-backed loans right through the Coinbase app, powered by Morpho, all in just a few clicks. So whether you're a builder looking to integrate Morpho or a user searching for the most flexible DeFi lending experience, it's time to check out the brand new app.

at app.morpho.org and experience the future of on-chain lending today. All right, welcome Bankless Nation. We are on the ground at Ethereum Denver. We'll call it AI Denver this week. I'm here with Ejaz, our regular AI Weekly Rollup co-host. This has taken the AI Rollup show...

slot of the week this week. And we're also joined by Tommy, who has been investing in the AI crypto space for a very long time. And we've wanted to do this episode for a while. Now we're doing it in person. Tommy, welcome to the AI Rollup. Yeah, thanks for having me. I listen every week and it's my favorite AI Rollup show. So proud to be here. Yeah, I think doing this episode with you has been a long time coming. Maybe I want to start off the conversation here. Delphi has been investing in the AI crypto space since I think even people really realized it was a space.

And so I kind of want to just go and roll back the years to whenever you guys started. What gave you conviction and belief in investing in this space in the first place? What was the thesis then and when was that?

Yeah, we started over two years ago at a time when we got significant pushback internally because we didn't, you know, some people just didn't view it as a viable space. Right. And that's how you know you're early. I got really interested through an existential crisis using mid journey and early chat. Right. Like this is going to take over the world. Are we in the matrix? How do I get exposure to this? And that

just drove me down a full rabbit hole learning curve of, I got to figure this out. I need to invest here. I got to be aligned. The best resource I had at the time was Stephen Wolfram's chat GPT explainer. It was incredible. One of the best threads. It was so, I appreciate it. Yeah. I did a thread on it. It, it oddly got like 3 million views or something. I pinned it and

The rest is history. We kept going. Did you have any sort of general idea about why it made sense to invest in that sector? Or was it just more of a sentiment, a vibe? Or did you actually have a clear thesis about what you were looking to see come out of this sector of crypto? Yeah. Initially, it was just, this is the new tech wave. This is the explosion. This is where it's going to happen forever.

Um, and I wanted the exposure to that over time, the thesis grew and cemented my view that I do not want to live in a world where Sam Altman controls a single model that controls our life. Right. And that is my North star on what I want to not, uh,

empower and what are the things I want to invest in? And it's pretty hard for people to understand what that means because people love using O1 Pro. They like using ChatGPT. They like getting their summaries. But when you have dating apps and insurance and banks and all the apps that you use today built on a model and it knows you better than you know yourself and it's able to convince you of things you don't even want to like,

a political belief, a girl, a guy, whatever, that is scary. So the open source crypto AI movement across the stack of AI is what I'm extremely interested in. How did you see crypto actually becoming relevant in that stack? Because you could have just pivoted and around the time you guys started investing in crypto AI was right around that time of the Jason Kalkanis, if you're in crypto, pivot to AI tweet.

So you could have just done AI and not done crypto. How did you see the crypto part of that ingredient mix in there? Yeah, I mean, two years ago, it was super contrarian, right? I mean, you looked at the big AI labs and you said, they have all the GPUs, they have all the talent, they have all the funding. There's no way we'll be able to compete with this. But the core metrics and tenets of crypto that we all know and love are

like rang true, right? Like instead of being in a small lab with a limited number of people with top down, this is how we want to build things. This is the alignment, the safety we want. You can have millions of developers around the world raise overnight for their startup, create a new AI application and change the world, right? We saw it with Uniswap. We saw it with DeFi. But the difference was we needed to know how to code with DeFi, right? In this world, you don't have to know how to code, right?

You can launch your token and build your startup with Claude or OpenAI or whatever. So the barrier to entry is literally on the ground. So I expect we'll see literally millions of agents and hopefully some product market. I actually have a quick follow-on question. When you discovered this match between crypto and AI, were there any burning uses that immediately kind of popped up? I'm curious.

I mean, the initial ones were the easy ones, like the deep ends around compute and your unused hardware and clustering. That was an easy one. But

And that sort of started our investments, right? Yeah, that was like the first real thing that people were like, yeah. Exactly. So we started investing at the bottom of the stack, like the infra players, the compute providers, the decentralized or distributed training projects. And we started moving up the stack to like middleware and inference, then eventually to like the app layer, like MyShell and things like that. But the interesting part is it's sort of playing out the opposite, right? We're seeing agents...

at the top having their virality, leveraging closed source models and infra and now backing into open source stuff. So it's working, it's just not the way I thought it would go. I think people who are in crypto but not necessarily in the AI side of crypto has probably absolutely noticed the Cambridge explosion of agents invading their Twitter replies. And that's kind of marked the AI crypto wave.

for them, unless you've been like actually what you've been doing, which is investing deep into the stack. Crypto AI just means like the last three months of virtuals, AI16Z, you know, Zerobro, some of these agents, and that has been crypto AI. But I think what most listeners who are, if that's been their experience, what they've been missing is like, there's actually been multiple waves

of crypto AI activity before the agent meta really blossomed at the top. And that's really the cosmetic layer of crypto AI. Maybe you can kind of run through the eras. I know it's only been since like 2022, 2023. But maybe you can kind of run through the eras that we've seen in investing in the AI crypto fundamentals since you guys started investing. Yeah, yeah, for sure. I mean, on the liquid side, BitTensor started the space, right? I have plenty of critiques, but they did start the space and they should...

they should be known for that. And that's on the liquid side. On the venture private deal side, it was a lot of those compute networks, inforplays, essentialized training, distributed training type investments. The likes of like Anoos Research or Prime Intellect, Ionet, Akash was liquid, that sort of realm. We spent a lot of time on the coordination networks for a while. There's Allura out there. There's Sentient, obviously. There's

There was Morpheus launched publicly, like a lot of trying to coordinate the model creator with the use case. And then we sort of went full fledged into the agent side and the model side. The models obviously power the agents, but the agents are what people see and use and love. So that's what we're seeing now.

I've always said that I think that the reply bot, infinite slot bots are going to zero. I've tweeted out many times over the last couple of months, and I honestly feel kind of bad that they did, right? It sucks, but...

it's not where the space is going, right? The space wants utility-based functional agents that are doing things, things that are trading, managing a DeFi protocol, managing risks, potentially creating bankless content. It doesn't matter, but agents that are doing things are valuable, and that's where I think we're going. And we're starting to see that, right?

There's a couple of agents out there, like Billy Betts is an interesting one, has 100K in its book, and it's betting its own on all NBA games, and it's powered by BitTensor subnet. So I think the move to the utility functional agent side is where we have to go. Mm-hmm.

Yeah. Is there any evolution or development in your own thesis about the crypto AI investment vertical as we've gotten more deep into this space? We're in 2025 now. We've seen agents come and then some of them, most of them go, but they're still here.

And so when you were just kind of like, you know, eating some cats and checks back in 2022, 2023, just kind of maybe on a vibe that, you know, AI is very cool and it's going to integrate with crypto somehow. Now we're in 2025. Has your thesis gotten like more precise or developed or iterated over the years at all? Yeah, it's interesting. Yeah.

Even six months or a year ago, there was significant pushback on the idea that we could compete and win against AI labs. Again, they have the GPUs and like the amount of GPUs that like Meta has is like nuts. Like there's no way, it's really hard to compete with that.

and the talent and their outputs and things like that. But you saw a huge inflection recently with deep seek, right? The idea that you have an open source model, that's a reasoning model. And that could search the internet before even anthropic had a searching or reasoning model out there is nuts. Anybody around the world could use that to create an agent, but more so it just demonstrates that the closed source lab could be, uh,

changed, right? Like there is real competition out there. And DeepSeq was expensive. There's hundreds of authors on that paper. It's a legitimate, huge innovation. But there are well-funded entities out there just like DeepSeq that can do something similar. They can take DeepSeq and change it and edit it and grow it. There'll be totally new open source models out there. And I'm sure OpenAI will get better and continue to lead. But

I don't necessarily think crypto has to beat open AI at what they're doing. I think we just need open source models that people could mess around with and create projects with. And that's how we win.

I think this is the big question, the trillion-dollar question, which is where does value accrue in the stack? And we're seeing this tug-of-war happen on the centralized AI lab side of things, where open AI maybe has the most sophisticated model, maybe it is generating the most margins for its model, but it's being materially compressed by people like Facebook Meta, which is very quickly open sourcing their not-sourced

they're inferior model, but if you just give it out for free and it's only 20%, 15% worse, then you're really compressing the value of open AI's lead. And, you know, meta is just giving this out to the open source world. And then the open source world, well, I've always interpreted that as crypto. And so like, where do you see, like maybe just like

how do you see the value accruing to either the centralized AI labs versus open source? And where do you see that equilibrium lie in terms of total value creation? And I see it as a tug of war between the open source side of the AI world and then the closed source AI world is always on the frontier. They're the best funded. They're going to make the most sophisticated models. But open source always seems to be so hot on the heels and is so seemingly able to monetize via crypto very well.

So like very broad, loose questions. It's just a broad stroke vibe of a question. But like, where do you see the tug of war between open source and closed source? And where do you see value accruing inside of those domains? Yeah, I'll get my answer. I'm curious your take afterward. But the...

The closed source labs are clearly winning and are clearly valued higher. They're worth hundreds of billions of dollars. Claude's worth, what, 60 to 80. OpenAI's near 200. Perplexity's like at 15. These are all extremely valuable, well-funded projects. I think that they will continue to lead on the foundational side in terms of metrics and what they produce. I don't think that's been changed yet. We have early signs of success.

News Research trained a 15 billion parameter model. Prime Intellects trained a 10 billion parameter model. They might have done other things. You're starting to see the world inflect, right? Because the last 20 years, data centers have been optimized for co-located GPUs and redundant fiber and power and gas lines and cooling and to optimize for that situation. We've only just begun to research and optimize around training

devices or training models with devices that are not co-located. So the ability to do two to three thousand times less communication per node, which is like what Noose is doing, is extremely important and is only gonna get better. That being said, on the open source side, I think it's pretty clear that they're gonna win on the app side. Because you just have the world's developers

iterating on consumer level applications and agents at a scale that I don't think apps can and my counterpoint to open AI is have you ever used the chat GPT app store and most people say no I didn't actually know this exactly so it's and but like we've all heard of agents for crypto guys but right yeah yeah yeah I actually largely agree with what you said Tom I think

The slight amendments I would make is I don't know if open source will win on the app player. And the only reason why I say that is

I think it'll come down to quality and curation of a bunch of these different things. And I also don't think it'll be an App Store-like experience for a bunch of these agents, right? I think you'll just have one kind of agent interface and then people just kind of like interact with it and it'll leverage and pull whatever service or maybe even other agents that it can, right? Like that's the whole like swarm model, right? And what that orchestration layer looks like, I have no idea. Will it be like a...

Google Play Store kind of situation where it's like Android or would it be a more curated fashion where it's like Apple gets to like pick which apps or services are like blue chip and then discards everything else, right? There's pros and cons for both, right? On the Apple side, it's like safety, security. You're not going to get rugged on the Google, you know, Android side of things. It's like, oh, I don't know if this is as good. And it's kind of like a lagging indicator. The other point I wanted to make is you said that

the centralized closed companies have the advantage because they have co-location and stuff, right? I would say another reason why they've got this lead is it's just data at the end of the day. I think a lot of these models will end up following similar techniques. I mean, look how quickly people replicated deep seeks learnings. And of course they open sourced everything, right? But then, you know, now reasoning models are popping up everywhere. Claude Sonnet 3.7 released this week and Grok used a bunch of it in,

in Grok 3. Grok's awesome. Grok's really awesome. Really good. And sorry, side note, I really like Grok 3 because it like leverages all the Twitter data as well, right? Yeah, it's real-time data. Dude, for Chrome.jp, it's just like, wait, what? Like, they just said this and it's not a fact? But I don't know, you know what I mean? So I love that part. And I think, so let's assume these models get, you know, replicated in some way, right? And let's say that reasoning, you know,

So this is what DeepSeq pioneered for the people that are listening, where it doesn't use as much data and it doesn't use as much compute. So it's, to your earlier point Tom, more accessible for other teams to compete with these top model creators.

I think it's going to come down to that data point. And I don't really know how open source gets ahead on the data side of things. There's a few ways where it could be like aggregating open source data, but I'm wondering whether personalized data becomes really more important. I think that's going to be like a key focus going forwards. Yeah. I mean, you brought up a lot of, I mean, my concerns are pretty similar, right?

in a world of millions of small, narrow models, which maybe they're trying to proprietary data, like the bankless data and the community trains a bot, which would be pretty cool to see, right? Or a Delphi bot or something like that.

I think that's key, like narrow models. And my data point I always use is Google wanted to trade a medical model two years ago called MedPom. It's my favorite example. And it was like 60, 65% accuracy. They gave a lot of doctors, doctors gave written responses to train the model and it got up to like 95% accuracy. Doesn't have to rationalize about the world, just a narrow segment. But to your concern, deciding what model to use and when is a huge unsolved problem, right? Like,

Do I trust the Bankless model or the Delphi model or the EJS model? Or do I use all three and average them? That is a really big unsolved problem. The other big unsolved problem that I think we're early on is the incentives and value around open source models. So if we are successful in training a trillion parameter open source model and the weights are public...

What does that mean for the token underlying it? Or is there no token? You don't want a bunch of random people just voting via the token. I love it now. What could go wrong? Love a good now. Yeah. But I think we'll figure it out. Actually, on the first point that you made, right, what do you think will...

push people to consider which model gets chosen, right? My guess is it's at the application layer, right? Let's assume all these models do the same cool stuff. Surely the winning kind of moat will be whatever application gathers the most amount of humans that's leveraging its AI capability to point to A, B, or C models.

Do you have a different view? I'm kind of curious. Oh, I mean, it's, I agree. It's early. But I mean, we're all, I'm already making mental decisions on models today, right? Like when I want to do, when I want like real time data on news or project, I'm using Grok. When I want something to think for 15 minutes, I'm using Owen Pro. When I don't want a,

30 page system prompt telling me how I should view the world. I use Eric Voorhees as Venice, right? Like, so I'm already, we're already making decisions on what model we want to use. The idea that robots would be like siloed to one model is just nuts. Like they're going to use way more models to us because they have way more, way bigger preference stack than we do.

I want to give a take your way and see how you like it or don't. We have all these models from the centralized AI labs. I kind of see them as different personalities. They have different skill sets. They have different skill points that have been allocated differently. And then maybe some part of the crypto AI or the open source AI is going head to head with these labs. And that's like the golden goose. That's the big, big valuable thing that's out there is generating the best model.

But there's also just a lot more long tail things that are out there. And we were talking before the show about some of these things that we even mentioned last week. The Billy Bet model who's just making bets on Polymarket, sports betting on Polymarket. Narrow application of a model.

And then there's the NBA announcer that I think is Tracy. I think that's really cool. And that is a narrow application of this model. And one of them is using PolyMarket, so it's using a crypto platform to do its job. One of them is tokenized, so there is a token. And so my attitude is there's like the fat tail, which is you're trying to make the world's best frontier model. And that's the fat tail of AI. That is AI.

And then there's like crypto AI, which is more of the longer tail, applied narrow focuses of these models to do certain things. And millions of people watch basketball, millions of people watch football. Maybe you want an AI commentator who's just the best commentator out there because they can do that. And you could make that not crypto. You could just totally make that not crypto. You could just make an AI commentator, stream that on Twitch, and that's the product.

Or you could issue a token and monetize that thing and have that be some $1 to $100 million token. And the internet markets of crypto allow for that to happen. And, you know, regulation dubious, but regulation aside, the monetization layer of the long tail of applied models seems to fit with crypto where, yeah, you have this

sports commentator named Tracy, and she's getting tens of thousands of views every single week. I don't think that's not true, but like theoretically that could happen. And then the monetization layer of crypto is where that where crypto innervates into that, like many thousands and thousands of different possible use cases of more narrow applications of these models. That's kind of how I'm thinking of the future of this space. How do you think about that?

Yeah, maybe just taking your example a step further, like, I don't know sports, so let me know where I messed this one up. I also don't know sports. EJ, maybe hold our crew. Nope. All right, the audience can make fun of us after. If you love... My wife loves Duke basketball, right? And...

a specific commentator will come on that you will like or dislike more, right? What if you can have your commentator tell you the game in the way that you specifically like? Maybe you're more interested in metrics, maybe you're interested in their life story, whatever. You can fine tune any type of announcer you want. I'm not saying that it's super valuable, but I'm saying that crypto is really efficient at figuring it out. Like the idea that all the reply bots

are down 90, 95% demonstrates that we are a efficient market. We went through a cycle where we thought something was really interesting. We found out quickly that, hey, maybe it's not. Let's move on to the next thing. And in my view, that makes me really bullish on crypto AI because we're able to go through the iterations of the capital markets, the innovation cycles way faster than anybody else can.

Hmm. Mm-hmm. Yes. What do you think about that structure? I might take the other side of that and say, so here's my view. I think that

tokens are very efficient in terms of incentivizing certain behaviors. I don't know if we've nailed the design, and the design varies from different use case to different use case, to make the right behaviors happen at the right particular time, right? And so when I think about, you know, just whether it's the example you just used or just tokens in general, I think number one, it's

giving people the funds necessary to build the thing in the first place. And I think it applies to AI way more than it applies to anything else because there is such a high cost barrier for anyone to play in this game, right? You know, you spoke about what got you involved in crypto in the first place. It was at the resource stack. It was at the computational layer. And back then that was pre-DeepSeek, right? People didn't have the funds to be able to play with that. That was BitEnser's first main pitch. It was like,

the stuff is expensive, but we have the incentives to be able to push different teams to build certain things. So I still think crypto's number one thing is on the fundraising side. And then when it comes to incentivizing efficient behaviors or the market being efficient, I still think that

There is a large weighting on narratives, especially when it comes to crypto AI. And that isn't a surprise, I think, to me at least, right? Because I don't think there's any formal recognition right now. And I know this is the boring, unsexy answer, but until we actually see a framework that recognizes these

instruments as some kind of like whatever pseudo equity stake or governance, right? Or whatever. People aren't going to take these things as seriously as I would hope them to right now. But we're seeing that evolution happen in real time, which is just pretty cool. I think you're right. A lot of the tokens shouldn't exist.

I think you're totally right on that. And I'm totally with you that the slot bots need to die, and they are dying. They aren't dying fast enough because I still have to block 14 of them a day. You have a huge following. You have every bot. I was about to say. Okay, so I want to talk about the line that separates Silicon Valley...

AI from crypto AI, because there does their crypto crypto's brand might actually never have been worse. Actually, maybe maybe after FTX, our brand was pretty bad. But right now, our brand publicly facing is not great. Mainly just maybe that's effective, the meme coin mania. But I don't think we're necessarily totally taken seriously as an industry. And and then there's the Silicon Valley side of the world, which is just in this absolute massive arms race where people are just heads down, focused on building the best. It's day and night. It's day and night.

Stay in that. So how do we close that gap? Why is that gap important to close? How do we close that gap? And maybe we can even just maybe try to add a little bit more color as to like how big or small that gap actually is. Do you have any indications on this, Tommy? Yeah.

Did I mention the Wall Street thing yet or no? No. Okay, so up this alley. So me and my partners all worked on Wall Street seven, eight years ago before Delphi. We left for crypto and we were laughed at. There's no DCFs for this. There's no cash flows. And now today you have BlackRock buying crypto faster than anybody could have possibly ever imagined. Hopefully more with price action today. Yeah.

The same thing will have to happen with San Fran and the open source world. One of the biggest risks to crypto AI, in my opinion, is not pulling in traditional AI developers from San Francisco. And it's hard, right? But I think that there's a couple of reasons why they will come over. And this is an exhaustive list. I'm curious what your guys' list would be, but...

One of them is just the freedom, right? Being able to run that model simulation or to try that experiment instead of this top-down directive on what you can and can't do. Breaking out of alignment, right? About what we want the models to do or to not do. The ability to fundraise overnight, to network with developers around the world without having to only work within your core group. It's all the core tenets of closed-verse open-source development.

all over again. And I think eventually that they'll like it, but I don't think that they like it right now. They'll like it because they're getting paid though. That's one of the main differences, I think. They're getting paid for the trad AI stuff? No, well, they're getting paid to bring that trad AI skills to the open source movement, right? Because typically, I mean, I don't know if you read, I think it was Andrej Karpathy's blog post or it was someone else, but

He kind of went into the depths of why DeepSeek is so important. And then like the whole point of like why, like he's a massive open source proponent, whoever this author was specifically. And I wish I was remembering his name right now. But he said the main kind of,

uh challenge that it faced was that there was just no money in it yeah and so you know to raise the funds to be able like everyone wants to do the good thing but it's kind of like if i'm not getting paid for it i'm just going to stay with my you know open ai job super comfy no total comp and work five hours a day exactly but you know it's funny like as soon as people started seeing you know eliza go to a crazy evaluation or whatever there were i saw a lot of people from like silicon valley kind of like poke their heads over the fence and be like hey what is this right

Oh, it's a GPT wrapper? I could do way more than this if I do A, B, and C. And you start seeing them getting involved in the discords and stuff. So I feel like that inertia might push them over, maybe. There's also just some really cool stories of what open source enables. And we've hit on DeepSeq a lot, but it's a good example for, I guess, one good and one bad reason. The good reason is when you ping a model like a giant billion parameter or 100 billion parameter,

parameter model, you're pinging the entire model, right? But with DeepSeq, since they don't have access to all these Nvidia GPUs, they had to get super creative, right? And so when you ping DeepSeq, you're only turning on a small portion of the model, 10, 20 billion parameters of the 671. And that's like necessity is the mother of invention, right? And that's the sort of stuff you get with open source. The flip side is, I think they monetize it through their hedge fund, right? So that's the counter is...

hey, we're going to release this. We know NVIDIA is going to go down $600 billion. Maybe we should take a short out. I don't know if they did or didn't do that, but I would be surprised if they didn't. Right.

Right, right, right, right. Yeah. Going back to the conversation on the Silicon Valley AI devs, because we want them in crypto. We want them to join the open source army, which I think is equivalent to crypto AI, more or less. Right now, I'm assuming paychecks in Silicon Valley for AI developers are as high as they've ever been. And the race, the AI arms race is maybe accelerating. Maybe it's just, it is hot currently. It's been hot. It's going to stay hot.

Right now, the returns on AI skills is very heavily, I think, on the AI side of things. And you match that with the negative sentiment in crypto land at the moment. Like right now, it's probably pretty cushy to be working for OpenAI, Grok, Meta, any of the AI, Google, Gemini. As soon as that S-curve of the AI arms race maybe starts to teeter out and becomes a little bit more flat...

then maybe devs can start to look elsewhere. It's like, okay, where are the returns? Where can I get returns? And maybe that's where open source AI starts to catch up a little bit when the AR arms race cools down. Just a quick hot take. How do you guys feel about that? Where are we on the S-curve, David? I have no clue. I have no clue. But I think we are accelerating. David, a specific point. Every time I think we're at the top of the S-curve, like, oh, okay, the models are kind of getting commoditized. Ah, GPUs.

a deep seek drops and I'm like, oh no, we're down here now. The last two months of model sophistication, I think I've been the fastest two months of sophistication growth I've ever seen. Yeah. And I was working with, I always go to ChachiBT 4.0.

And I was asking questions about like Celestia as it relates to like settlement versus DA and consensus. And I was asking it like crypto jargon terms about like, okay, if a rollup is on Celestia using Celestia for DA but not consensus, what does that mean for property rights on Celestia? These are very deep crypto questions. Very specific. Very specific.

It was giving me very sophisticated answers. You can only love a model for a month, right? Like, I loved 01, now I like rock, now I like 3.7 on cloth. Like, you get a month, I think, maybe less, and then you got to move on. So, like, I think my answer to the question of, like, where are we on that S-curve is, like, we are on the acceleration of the S-curve. We are getting steeper, not shallowing out. But it's interesting, though, because, like,

Every founder, I think, other than Sam, has left OpenAI. They've all left OpenAI. Like a big Ethereum co-founder. Exactly. If you knew that this was the leading company and you co-founded it, why wouldn't you just stay there and milk that? There are a lot of reasons why I think people leave. The other thing is...

I love O1 Pro. I love O3 Mini High. I love that it could search. I love the deep research, but I still use Grok every day. I still now love 3.7. I think there was a significant moat in OpenAI for a long time. I think that they attempted to do a regulatory moat type situation going to the government with that, and they did.

but it's pretty clear that I don't know if the moat is as strong as people may think right now. - You've got Microsoft pulling out of the data centers now. - But SoftBank covered the tab. - Oh, wow. - So it's not a good sign for sure, but yeah. - Wow. - Introducing Unichain. Built for DeFi and powered by Uniswap, Unichain is the fast decentralized layer two designed to tackle blockchain speed and cost challenges. With this mainnet now live, you can enjoy transactions at up to 95% cheaper

than the ETH Layer 1, all while benefiting from an impressive one second block time that will be getting even faster very soon. Unichain is the first Layer 2 to launch as a Stage 1 roll-up on Day 1. That means it comes with a fully functional permissionless proof system from the start, increasing transparency and further decentralizing the chain.

More than 80 apps are joining the Unichain community, including Coinbase, Circle, Lido, Morpho, and Uniswap. You'll be able to bridge, swap, borrow, lend, and launch new assets and more from day one. Built by Uniswap Labs, the team behind the protocol that's processed over $2.75 trillion in all-time volume with zero hacks, Unichain truly enhances DeFi experiences with faster, cheaper, and seamless transactions, even across chains. And soon, the Unichain Validation Network will allow anyone to run a node and earn by securing the network.

Visit Uniswap.org and swap on Unichain today. Ondo Finance is leading the way in the tokenization of real world assets with a mission to accelerate the transition to an open economy by building the platforms, assets, and infrastructure that bring financial markets on chain. With Ondo, institutions get seamless access to tokenized finance while on-chain users gain access to institutional grade real world assets like never before. By bridging traditional finance and the blockchain economy,

Ondo is unlocking liquidity, efficiency, and global access, reshaping markets for the digital age. As the leader in institutional adoption, Ondo finances the gateway to the trillion dollar tokenization revolution. Institutional great finance, on-chain, for everyone.

Celo is transitioning from a mobile first EVM compatible layer one blockchain to a high performance Ethereum layer two built on OP stack with Eigen DA and one block finality all happening soon with a hard fork with over 600 million total transactions, 12 million weekly transactions and 750,000 daily active users. Celo's meteoric rise would place it among one of the top layer twos built for the real world and optimized for fast, low cost global payments.

As the home of the stablecoins, Celo hosts 13 native stablecoins across seven different currencies, including native USDT on Opera MiniPay and with over 4 million users in Africa alone. In November, stablecoin volumes hit $6.8 billion, made for seamless on-chain FX trading. Plus, users can pay gas with ERC-20 tokens like USDT and USDC and send crypto to phone numbers in seconds. But why should you care about Celo's transition to a Layer 2? Layer 2's Unify Ethereum, L1's

fragmented. By becoming a layer two, Celo leads the way for other EVM compatible layer ones to follow. Follow Celo on X and witness the great Celo happening where Celo cuts its inflation in half as it enters its layer two era and continuing its environmental leadership. I think that the story of all of the OpenAI co-founders leaving to create their own OpenAI competitor is illustrative of how incredibly valuable

Silicon Valley AI labs are. 100%. That, to me, tells me we're still on the very early, because like, Mira just launched, what was it? Thinking Machines. Thinking Machines, and then Ilya. Yeah. And those are just, like, we haven't even seen their models yet. And those are just brand new entrants. Mm-hmm.

And so if you make this into an Ethereum parable, Avalanche just got made using the EV. And so now we're just going to have... Maybe you haven't even seen the Solana to Ethereum. We're seeing the fracturing of the Layer 1. We're seeing the Layer 1 trade kind of show up in... Or the properties of it show up in the AI labs. Now there's like seven pretty high-quality frontier model lab teams all making very strong models. Yeah, I don't think that we're...

I don't think crypto AI is by any means far and away the winner. Right? Yes. I'm here because I think it has outsized returns. It aligns with my worldview. And I think that the capital raise mixed with the creativity of AI leads to an explosion of applications and use cases. I think that's right. And that in turn draws in more of your traditional AI engineers too. Exactly. Yeah. Yeah. I kind of see it.

Kind of going back to the question I brought up earlier is I kind of see crypto AI as the long tail, the periphery of everything. What did Uniswap do really, really well? It provided liquidity to things not Bitcoin, not ETH, but the long tail of assets.

And so the blue chip of AI is going to stay inside of Silicon Valley. It'll go public. OpenAI maybe goes public one day. But these are the public companies. But then the crypto AI side of things is really this long tail kind of crazy expressivity where you get out of your nine to five trad job and then you go into the wild west of crypto. And there's so many more possibilities that you saw out there. And there's like 1 million people

10 million dollar companies that you can tokenize in crypto. Whereas on the Silicon Valley side, there's 10, 1 billion, 10 billion dollar companies. That's kind of how I see it. I mean, I agree with you. What I would add is

I think people tend to overlook the scale of what crypto can enable, particularly on the AI side of things. And I think a lot of reason that people think that way is PTSD from scaling blockchains themselves. It's an absolute pain in the ass and no one's going to deny that. But when you put it in the kind of

perspective of AI, it's like, okay, so let's create a bunch of these AI models. Let's create access to it via compute, inference, et cetera. Okay, so we need chip cost, all that kind of stuff. Then it's like, okay, now let's allow a bunch of people to use it. What's the easiest way to use it? Oh, well, we have chat boxes. What's a better way to use it? Agents. Okay, the agents will allow me to do a very specific thing. Book travel, you know, arrangements and stuff. But then it's like, oh,

they could be so much better. They could hit A, B, and C parks. What if we brought in A, B, and C? And then we could give it a wallet. We could give it a bank account, guys. And then suddenly you run into a million different errors, right? 100%. Error 404, Chase Bank is like, I don't know if I want to support this. Meanwhile, he's buying the bags behind the scenes. Do you know what I mean? So I think crypto being open source, I know it's so...

cringy and lame is actually one of its moats that I think. So it's not only the long tail. The long tail itself is like one of the things. It's that...

The fact that you can just orchestrate on this open play is like a canvas, basically. And I think every day we're getting closer and closer with every L2 that drops. Closer and closer to economies of scale for transaction across or a microtransaction or an API inference or whatever that might be.

And that's what really excites me. Are you kind of alluding to the money Legos for agents? Yes. The AI side of things? Yes. Money Legos for agents. Agent Legos. I truly believe that one of the top app services layer for agents will exist in the crypto world. I'm not saying it'll be a separate L1. Maybe it will. I have no idea. It's too early. Right. But I believe some kind of middleware layer will definitely exist within crypto and it'll be the go-to method. Yeah.

I can't see how everyone just decides to kumbaya and work together for the greater good. I don't see that working out. Okay, so as we are wrapping up this AI agent slot bot meta on Twitter, and as I want to really emphasize to the audience, we've seen AI crypto bubbles come and go before. Maybe bubbles isn't the right word, but just waves of investment or themes of investment, maybe two or three of them in the last two or three years.

Um, so now, now we're, you know, halftime or break, there's a break and then there's a lull in the excitement. There's a lull in the bubble. The idea is that it comes back again. Uh, that's why we're investing in it. That's why we're making content in this space. Um, so maybe we can kind of just talk about what we're excited about in the future. Um, what themes that we're looking forward to, um, maybe Tommy, you can share any exciting checks that you guys have wrote at Delphi about what excites you. Uh, maybe actually let's start there about like what excites you about the frontier of AI crypto. Yeah. Um,

We've made 20 crypto investments last two years. Okay across the whole stack So it would be hard to call them all out because I can't but I'm really excited about the distributed training side the news researches of the world I really like that. I'm really excited to see Agents that use reasoning models things like that from a personal account perspective. I'm really excited on the framework side. Mmm

I think AI16Z and Arc are really interesting here. They're open source, hundreds of plugins, a lot of developers. I hope I don't get shot for saying I think Sol's doing pretty well. I say it every week. Yeah, okay, cool. All right, I'm free there. But Base is also doing really well too. But I think that the...

The reason I like the frameworks on Sol a little bit is because they're open source. They're leaning into max developers, which is where the creativity comes from, I think. And I really like that side of it. On the base side, I think we've seen great new agents from virtuals. Like...

The betting one we've been talking about a lot. And I think Venice is... Sorry, we've seen Venice too. But I think virtual is also fully leaning into the tech side, which is nice. They launched their agentic commerce protocol. I'm going to mess the name up with the lemonade demo, which is really cool with all the agents working together. And they also are fully leaning into game, which is nice to see too. So yeah, I think my thesis for AI is...

to take from Vishal from Mythos VC who we invested in is you need a proprietary model, data, or business integration. And now we can add on

maybe specific proprietary tooling from the frameworks or changes to the reasoning model and inference time compute. Those are the five sort of areas I'm looking for. Can you actually unpack those? The first three, because you gave me those in a call one time. I thought that was really useful. The model, data, what was the other one? The business use case? Business integration. Can you talk about just those three ways of thinking about a crypto AI product? Yeah, so...

differentiated proprietary model would be, you know, a model trained on all of the bankless data or Citadel trains a specific model trained on all their data, right? That's a specific model. It could be a new architecture like deep seek or a fine tuned version of Lama, right? Uh, the specific data is what he jazz alluded to as being so important, right? That data to train a use case that's not on the internet already, uh, could be very valuable for the end model, which would fine tune or the business integration. So think like, um,

you know, somebody, some chatbot gets a specific integration with Uniswap, they have full distribution or the agent that works with Google sheets has full distribution. The two new ones I think are the framework side. So the specific new unique tooling that you allow agents to do, maybe that's creating a wallet, posting social media, whatever they can offer. And then the last one is the inference time compute, you know, thinking for longer or thinking in different pathways. Um,

I think are gonna get pretty interesting. 'Cause when you get inference time compute, if I'm asking a math question, I want it to think in math terms and make the right directional choices and lefts and rights on the street. I don't want it thinking about philosophy and things like that. - So when you're just looking at a deal at Delphi that's in the crypto AI sector or the open source AI sector, this is kind of your framework for understanding your interest level

And if it doesn't have any of these things, then you're like, pass. And if it has maybe one or two of these things, then you get peaked. Yeah, we get a lot of deals. And I pass pretty quickly if the founders don't have real AI experience at this point. Because the space, even our crypto portfolio companies pivot a lot. And I think that's fine for AI projects, given how fast the space moves. But it becomes even more of a founder bet. Because you not only have to pivot, you have to pivot as a normal business, but now you have to pivot online.

on the scale of the AI changes, which is nuts. So they really become founder bets. So real AI pedigree, maybe experience out of previous AI lab has built something differentiated that we can use and see, or just a first principles approach to AI.

changing how we're looking at the world. And I think those are the biggest bets that we could find. - The Self-AI agent is like your guys' intern. So you guys just agentified your intern. Do you ever process a deal that doesn't go through the agent or do you just funnel everything through the agent?

I funnel a lot through the agent. I even funnel our investment committee docs sometimes to see what it thinks about the analyst pitches. But it picks up a lot of stuff that is really important. One of the biggest things it picks up is, so we have Claude 3.7 doing the analysis based on a really long prompt and scoring methodology that all Delphi worked on. But the most useful part is we added a step to send the founders to perplexity to do reference checks and find any red flags. Oh, wow.

So one issue we found was we would send it to Perplexity, but it would give us the wrong David Hoffman or the wrong E-Jazz, right? So the Perplexity search has to match two of the pieces of information from the deck. So you went to Harvard and you worked at...

Morgan Stanley if it if it matches that it'll return the post and it's returned red flags on founders fraud really prison all types of things before you'd be shocked what you find when you just search That's crazy yeah, and do you have plans to enhance it in any way or are you just kind of keeping it as is right now? Um

Sarah RGC would kill me if I started a venture funder on it, for sure. But she's awesome. But I want to make it a research model. I want a reasoning model to it. My partners, Jose and Jan, mentioned that it would be great if the founders can talk to it, add more information, ask it questions. When Owen Pro asks you questions, when you respond, they could ask the founders and then give us more info. That would be cool if it was a two-way street.

But I mean, I'm looking for somebody. If somebody wants to take it to the next level, I'm always interested. Have you ever, okay, so maybe the most valuable thing about this bot that you've created, this agent that you've created, is that other VCs don't have it. And that's why you guys want to keep it proprietary. But just hypothetically, if you guys wanted to monetize this agent, how would you think about going about and doing that?

That's a great question. It's hard, right? The monetization for us is the VC upsides and the deals at sources, right? Saves you time. Exactly. The other issue is like if we open source it and everyone sees the prompt. You lose your edge. You lose your edge, but you could game the prompt. Like if it says like I'm looking for, you know, really good looking founders and you guys apply, it's like I'm screwed. You know what I'm saying?

So it's hard to maintain that moat. Interesting. All right, guys, I think this has been really great. Maybe I just have a few questions that are just completely selfish. I use OpenAI 4.0 because...

I don't know why. And I know that there's other, like the Bankless engineering team loves Sonnet 3.5. Sometimes I use Perplexity. I don't know why. In your own personal life, in your own personal stack about how you use these agents, what agents do you like? Or what scenarios would you use a different agent? Or what's your personal toolkit for AI enhancing your life? How do you think about this?

Yeah, deep strategic thinking questions. I love 01 Pro on or Deep Seed, either one. For coding, which I obviously don't know how to code, but it codes for me, I would use Claude or for AI Bot. For real-time information, Croc all day long. The moat of having real-time user data is nuts. All the news happens there, all the top stuff, so it's great.

And then I really did like perplexity for a while for the online search, but that has sort of eroded quite a bit. For me, I use, so like you, I don't know how to code, but that's where Cloud comes in. And I think it's like,

We're developers now. Yeah. Yeah. Yeah. Yeah. Once it's vibe coded, I'm a vibe coder right now. And that's your resume. Yeah. Yeah. So cold was already pretty good at this stuff, but now with 3.7 and their new coding tool, do you remember the name? Versa?

No, no, no. They have like, it's like the Claude Agent coding tool. They released it synonymously with 3.7. I can't remember the name. It's on that list. You know what I'm talking about. It's really, really amazing. But then Cursor also that you mentioned is like just super cool. It literally is vibe coding. It's like, hey, I want to build this app. And this is the reason why I like Claude

in particular, and it just builds it kind of like life right in front of you next year. And you're like, oh, I don't like the color of the background of this game that I just created. Can you change it? Yeah. And can you make the size of the wall bigger and all this kind of stuff? And I'm a very visual learner. And I've spoken to my software engineer friends, and they're like, the number one thing is

to have real-time feedback and translation as to what I'm writing, but also to have contextual awareness of the code that I'm writing is just a complete game changer. So I use Cloud for that. I use Grok purely out of entertainment purposes because I spend an inordinate amount of time on Twitter. And to have like a real-time search analysis on Twitter, which I think honestly has one of the most valuable and also atrocious databases, but also one of the most valuable databases,

It's just like a complete unlock for me. And then ChatGPT, I'll be honest with you guys, knows more about me than my entire family combined. Wow. And the number one reason for this, I think I mentioned this on a previous episode, David, is I use voice mode. Right.

Oh yeah, that's an unlock. And when I use voice mode, you walk around and talk to it? Guys, it sounds like a human. Yeah, and there's also no delay. It's amazing. So I'm like, oh my god. And then you're giving really good advice, John. You're giving me really good advice, John. It's also crazy because you say please and thank you. You don't have to.

Yeah, but when they eventually take over the world, hopefully. Yeah, you're right. I'm actually a big fan of this thesis. My long-term idea about AI is that at one point it just becomes sentient. It becomes alive, very much equivalent to how humans are alive. There is no line that is crossed when we get there, so you might as well start being polite to AI. Yes, 100%. So you give them your please and thank yous. Yeah, you give them your please and thank yous. We will give David a larger jail cell because he said thank you.

More compute credits for you. But yeah, no, that's where I like, those are kind of like the buckets that I spend my time with these models. And I think like, I agree with you, the perplexity thing kind of, I don't want to say like fell off because knowing the rate of change now, next week you're going to drop something pretty amazing. And to your earlier point, Tom, like literally there's a new model dropped every week. So I have no idea like what's going to change by the time we air this episode.

Well, guys, this has been a fantastic episode. Before we go, we are still in the meat space and it's somebody's birthday today. And so we are going to all celebrate each other's birthday with this. Happy birthday, man. You're about to like splat that on my face there. Yeah, that's right.

Here you go, EJS. Happy birthday. Thank you so much. I think this is our 13th AI roll-up, so happy birthday, brother. Thank you so much. I pray that the agent market cap goes up only for the next two years. Amen. Happy birthday, man. Thank you so much. That was awesome. All right. Bankless Nation, thank you to the Soji House for hosting us. It's been great. David, thanks for having me. Tommy, thanks for coming on as a guest. I really appreciate it. EJS, I hope you enjoy your carrot cake. How is it? Razor at a time. Agents can't enjoy this.

All right. Bankless, hey, should you guys know the deal? Crypto is risky. Crypto AI is even riskier. You can lose what you put in, but nonetheless, we are heading west. This is Frontier. It's not for everyone, but we are glad you're with us on Bankless Journey. Thanks a lot. Thank you.