The Decade of Data (Tomasz Tunguz)

What's really exciting for crypto is you have the entire U.S. stock market, which is the largest stock market in the world, and then you have crypto, which is large, but not as large. And all of a sudden, those pools of capital should fuse, which is great. I think software startups will have a much faster path to IPO as a result of crypto than they've had in the last 10 years. How does this change end up accelerating the timeline to IPOs for some of these software companies, do you think? We have made it so expensive to go public that if you're a micro-cap software company, if you're worth somewhere between one and a half to two billion, not economic to go public. Most of those companies get taken out by PE right away. And so wouldn't it be amazing if you had a venue where small cap, high growth software companies were publicly traded? Hey, I'm Mario, and this is The Generalist Podcast. As the saying goes, the future's already here. It's just not evenly distributed. On this show, I sit down with the founders, investors, and thinkers who are living in the future. to help you see it earlier, understand it better, and capitalize on it. Today, I'm speaking with Tomasz Tungus, the founder and managing partner of Theory Ventures. In just two years, Theory has raised nearly $700 million in capital and backed promising startups across AI, crypto, data analytics, and data infrastructure. As Tomasz shared, his goal with Theory isn't just to build another venture for him, but something closer to an investing corporation. In our conversation, we discuss... what it takes to build a modern venture firm, and how theory is accelerating deal flow analysis and market mapping with AI. Ethereum's existential moment, how the blockchain risks losing its dominance the same way that AWS lost ground to Microsoft Azure as the result of AI. And the 100-agent future of work, why knowledge workers might soon manage an army of AI agents each. I walked away from this conversation with new ideas on how AI might reshape work. how crypto is changing IPOs, and why we're living through what Tomáš calls the decade of data. This is a new podcast, so if you like it, I hope you'll consider subscribing and joining us for some of the incredible episodes we have coming up. Now, here's my conversation with Tomáš. This episode is brought to you by Brex. Fred Adler, the influential venture capitalist of the 1970s, was known for displaying decorative pillows in his office that featured a signature business philosophy.

Corporate happiness is positive cash flow. In today's post-SERP environment, Adler's wisdom feels particularly relevant as founders need to make every dollar work harder. That's exactly what Brex delivers. Their modern finance platform was built specifically for startups like yours and designed to help extend your runway when capital efficiency matters most. With Brex, you get global corporate cards with up to 20x higher credit limits and no personal guarantee required. Their banking solution has no minimums and no transaction fees, while letting you earn high yield from day one with same-day liquidity. Best of all, Brex knows you were born to build, not juggle spreadsheets and finance tools. Their AI-powered platform brings cards, banking, expense management, and travel all in one place. It's simple, scalable, and designed to get you back to what you do best, building. More than 30,000 companies, including one in three U.S. venture-backed startups, trust Brex to help make every dollar count toward their mission. Join them at brex.com slash Mario. All right, Tomáš, it's so good to have you here. I've been an admirer of your writing and your public thinking basically since I first learned about what venture capital... was, which is now almost a decade ago. And so, yeah, it's great to have you here. I'm super excited. And maybe we can begin with a little bit of an introduction to your career in venture. How did you end up here and running Theory Ventures? Yeah, when I was 17, I started a company with my dad. It got me into startups. I was fascinated that you could start a company. And then after college, I studied mechanical engineering and computer science. And then I went to work for a startup that was founded by three alums from my school that ultimately ended up going public and was an early employee there and fell in love with it more, understood a bit more about it. Went to work at Google, built some large-scale machine learning systems with a phenomenal team for ad targeting. And then I've been a venture capitalist ever since. And I remember I was walking along the Embarcadero, which is the...

a part of San Francisco that butts the water. And I saw an angel investor talking to a founder outside at a restaurant. I was like, oh my gosh, I can't believe that's a career. I hope I get to do that one day. And I was very lucky that I had the opportunity to start and now have a firm of our own called Theory. And Theory, I think, is two and a half years old. Is that about right? Exactly right. Yeah. Okay. And one of the things that I really admire and find really interesting about your approach to investing is you both invest in data companies, but also take a very data and thesis-driven approach to venture, which especially at the earlier stages isn't particularly common, especially not in the way that you guys do it. How did you land on that approach as being the right one for you in the place where you felt you had a real edge? Yeah, I think there are many different strategies that make money in venture. And the one that appeals to me the most is to perform really deep research. And that's the kind of work I really enjoy. It's gratifying because it helps us build, as a firm, it helps us build networks of people who might be helpful to portfolio companies later. It helps us understand what's happening within those ecosystems so we can be helpful board members. And over time, we think that knowledge is a compounding advantage. The nuances of one wave, if you have them. in a particular place stored, let's say, in the organizational memory is a huge advantage later on. Insight Partners is famous for this. They have a database of all the pitch decks and metrics over their entire existence. And that was a big inspiration. And I look at some of the hedge funds that have done really well. They build information asymmetry with compounds over time. I think the same thing is true in venture. The data sets are obviously quite different, but if we build the internal systems of knowledge the right way, then I think over long periods of time, it compounds. That's really interesting. I would say that one of the maybe more common ways of thinking about venture would be that these asymmetries get sanded away super fast because things just get competed away really, really quickly. What are the types of asymmetries that you see as actually being compounding or having the ability to compound for you guys? The real asymmetry is just getting there earlier.

I mean, the venture capital asset class has grown from 8 to 250 billion in 15 years, let's say. And the amount of competition has increased dramatically. The number of different financial products that are offered to founders has grown. And that's wonderful for the ecosystem. I think the bigger startup land, broadly speaking, is the better. The goal for us is to be able to carve out a niche. It doesn't have to be very large, but it has to be one where we think we can create a great business and drive. strong multiples. And so to be able to do that, what we do is someone comes in on Monday, and I'll give you a story from our first intern. His name is Alex, who went on to start a company. And Alex, a brilliant guy, walks in on Monday during the investment committee and comes in with a 10-page paper. And he says, I don't believe GPUs are the future of AI. It'll all be ASICs, application-specific integrated circuits, which are chips that are designed to perform one operation like Bitcoin mining. But in this case, transformer architectures. Now there are startups that are doing that. And so he walks in that Monday and we debate. And I think that kind of investing is really fun. It's really fun because for half an hour every Monday, we sit there and... And within our team, we have a head of AI who is head of AI, three unicorns. And there's Lauren who built a healthcare practice at Palantir and sold contracts worth tens of millions of dollars to the largest healthcare companies in the world. And then we have two sales leaders from unicorns and then we have investment. So there's this composite team with many different backgrounds and experiences and they each have a different perspective on the opportunity. And that composite view and that initial debate really leads us to decide. Should we go and invest time here? Is this an investable opportunity? Is this within our scope? Does it meet our requirements? And that kind of investing is fun because out of that Monday, there's a lot of energy, right? If we find a theme or an idea that we're really interested in, there's a lot of energy. The whole firm pulls its resources, its network, and we can go and start a diligence process and a market map and really understand that space. And we follow up on those market maps every Monday. And then every month we review.

And we present to each other what we've learned. And so there is both a product for the organization, which is all of that research that we keep. And then there's an education of the entire firm. And so with time, we should all become increasingly sophisticated investors or increasingly sophisticated on many different spaces. So many interesting threads that I want to dive into there. One of which almost a bit of a meta point, but it's interesting to me that your investment team sounds like so many of them are former operators have really been in the weeds in industry. How did you think about building out the team that way and having these more functional experts versus folks that maybe have longer investment track records? Why is that the right approach for this style of firm? looks an awful lot like the future of, they look like investing corporations. They don't look like partnerships. Why are venture capital firms partnerships? Well, it's because of our history. Silicon Valley started with six or seven men who would get together every Tuesday at a particular restaurant and each write a $50,000 check into a company. And that was the first syndicate, right? And so, and out of that, two or three of them got together and said, let's start a firm. We can raise some institutional capital. Capital markets were not very supportive, so the funds were small. And then there were 10 or 15 of these firms up until the late 90s. And then all of a sudden, there was kind of this institutionalization of the asset class post-GFC, post the global financial crisis. And the asset class started to explode, right? But we have kept this notion of a partnership. And there are many different reasons for that. But I think ultimately... If you look at hedge funds, they're structured as investing corporations. If you look at private equity firms, they have investing teams and operations teams. And if you look at most major financial companies, there's a marketing team and there's a product team or technology team. And so we think that's the future. And in addition to that, it leads to better investment decisions because you have these different perspectives. We want our sales leader's perspective on

How effective a founder is it's selling? We want our head of AI and our technical team's perspective on the technical depth of a company. And we need those conversations and that understanding to happen very quickly. We can't have a lot of additional meetings. Ideally, we help the companies and give them guidance if we can. And we have a perspective on their market or their sales strategy. And it helps inform our investment decision. And so I think it just gives us like a mosaic or a better triangulation of understanding a particular business or market. Super interesting. I'm also interested to dig in a little bit deeper onto this thesis development piece. you are constantly generating these theses as an organization and sort of having these Monday meetings and chasing down new leads. But also you've sort of organized the firm a bit around sort of a few mega theses that you are deciding to like really go deep on sort of data, AI, crypto. You have much sharper lenses around those different things. How did you sort of land on those as, let's say, the sort of super narratives that you're following? And how do you sort of convince yourself that these are the places where we really have to go deep so you don't get stuck chasing a dead end or chasing something that's so, so broad that you would really actually end up having no edge? All of them are underpinned by data. Modern data stack, databases, visualization, that's data. AI, machine learning, all of that's data. And blockchains, well, that's just a different kind of database. Right. So they're just data systems broadly written, and they are sold to different buyers. That's really interesting. So then what is the deep interest in data systems there? Why is that the thing that you're like, this is the big opportunity for us? Well, I love data. Yes. Part of it is personal, right? That's a prerequisite. So number one is I love data. I think the second thing that's, they create really big companies. I mean, Databricks and Snowflake and the modern data stack are both Decacorns. Maybe we'll see them becoming Centacorns. Within the world of AI, you have Anthropic and OpenAI and many others that will be crowned Decacorns. And then within the world of blockchains, I mean, you have Ethereum is worth 350 billion. Solana is worth 90 billion. So you have incredibly large outcomes. That's one.

Two, there's a combination. So all of these markets have technical innovation, faster databases, in-memory databases, better forms of decentralization, new AI architectures. Will it be transformers? Will it be diffusion models? Are they state-based models? And so there's a technical component that is fun to understand. And then there's a go-to-market component, right? And so that's fascinating. And it filters, right? There's less competition in very technical domains. just because it requires a particular set of expertise and a desire to understand it. And so the combination of those two things, really, really large markets, the sort of barrier to entry for complex topics. And then the last thing I'd say is all of these categories are of the moment, right? They're very interesting categories. They're capital efficient. And so that makes it fun. Plus the replacement cycles are fast. So in venture capital, we might invest in a company and work with that company for five, 10, maybe 15 years. But ideally the cycles, it's not like nuclear reactors, right? Let's say awesome space, but a nuclear reactor takes 10 years to build. And then the next generation is probably 50 to 60 years from now. So you might be able to invest in one nuclear reactor company in your lifetime, but you can invest in many data companies in one's investment career. I think to a certain extent, everyone's interests are sort of inscrutable even to themselves. Like it's hard for, you know, if someone asked me, why do I like writing? It would not be an easy question to answer. But to the extent that you can articulate it, like why do you love data? Like what is it that speaks to you about that as something to spend your time on? What is the elegance of these business models that like really animates something in you? Well, it's all understanding. I remember being in, so I had this amazing professor in grad school. His name was Ming Pham. And he, this is in a class, I studied control systems. And our two final exams, one was building the autopilot system for Boeing 747, given only six sensors. And then the other one, which was a real use case, was imagine you have to send, you have to receive a signal from a satellite and there's a pad that has six actuators and that six actuators move the receiver.

And you have to move that receiver in real time to maximize the signal given earthquakes and the earth moving. And so, you know, whatever, like cool technical problems. We learned very basic control systems. And I was just fascinated by the fact that we could take the sensor data and actually make something that was magical, right? Like a plane can fly by itself. I drive in a car that, well, I don't really drive the car anymore. The car drives itself. And all of that is as a result of a whole bunch of sensor data that's coming off some telemetry. The same thing, you know, he was doing something in the stock market. And so it's kind of like fracking. It's kind of like oil where you take this thing that comes out of the ground and then you can do everything. I mean, you can make gasoline to put in a car from oil. You can make plastics to have a Tupperware container to save your applesauce. or you can make clothes from oil, right? And so I think it's this unbelievably useful raw ingredient and we're getting better and better at it, right? The whole world of data before say 2010 was focused on the very finest grade of oil, Brent, let's say, which was structured data, stuff that looks like it's in Excel or could fit in a structured database. And then over the last say 15 years, we've been learning how to frack data. We've been taking this. raw unstructured data, the Word documents, the conversations that you and I have, the podcasts and blog posts. And we're able to run it through an extremely inefficient system, but actually get something useful out of it with AI. And this is why I think, or at least certainly one of the main reasons you wrote about the 2020s being the decade of data. What are sort of some of the other things that like make you feel like we were really at a potential inflection point for this magical material? Well, I mean, just look at the capex spending of Facebook or Meta and Google, $250 to $300 billion in data center spend. And you read the public earnings of all of those companies, they're capacity constrained. There's way more demand for AI consumption than they can supply. Meta is building one that's larger than Manhattan. One of them that's slated to be produced will consume more electricity than the country of Ireland. Wow. And so when you get a sense of this scale of the computers, the data centers, but really computers that we're building,

I was reading a research report. We think in five years, 15% of American electricity, which is growing, the demand, 15% will be used to feed data centers. That gets me goosebumps. I'm so excited because it just tells you how valuable all of these data pipelines are and the demand for those kinds of insights. Given what you said about your intern Alex, who was talking about things down to the chip level, do you look at companies in the hardware stack and down to the chip level? you would theoretically invest in or no, too far field? By exception, there's a very particular set of expertise that you need at the chip level. I mean, there's a whole field of EDA, which is CAD drawings for these chips. There's the assembly of chips on one side and power management on the other. It's a fascinating domain. I think the hard part, the couple of challenges, at least for a fund of our relatively small scale. The first is the amount of capital acquired for a lot of these chip companies is pretty significant. The second is if they make a mistake on tape out, which is kind of the final rendition, it's extremely expensive. And then the third is the number of potential buyers is ultimately pretty limited. I mean, look at the EDA market, the software market. There are two or three publicly traded companies. Synopsys is one. Cadence is another that could buy these kinds of companies. There will be breakouts, there is no doubt, but I think it's hard for us. And then the other dynamic there is you have all of the major hyperscalers investing in their own silicon. So you have Google with the TPUs, and that's not going away, and Amazon with the Inferentia chips, and I forget the name of their training chips, and Microsoft has theirs, and I'm sure Meta has created open source architectures that they've released. So I think it's a big boy's game. Yeah, hard to break in as an insurgent. On the crypto side, you've been interested in crypto for a long time and you really are interested in it for the sort of core technology rather than a lot of the mania that surrounds it. A decade, decade and a half in, depending on how you want to judge it, how do you think about the progress crypto has made as an industry? Where do you sort of see it in the cycle of development?

I think it's broadly underappreciated and it's been extraordinarily massive. I mean, you look at Bitcoin for a new generation is gold. It's the equivalent of gold. People trust it more. And that I think would have been, that's a pretty phenomenal statement to say we've had an asset that has existed for thousands of years in gold that people trust as an ultimate store of value. And now we have a digital equivalent. And I think it's, you know, less than 50 basis points of global wealth is in Bitcoin and most of the projections say it's going to two or three. So that's pretty significant. I think the second thing that's really happened within the world of crypto that we're seeing now is the broad acceptance within the U.S. of stable coins. Yeah, yeah. Absolutely. I mean, by the end of the year, I bet most major large banks in the U.S. have their own stable coin project, a pretty significant fraction, five to 10% of. of us dollars moving through systems will likely be through stables and i think two years ago that would have been unfathomable yeah that seems like it's happened insanely fast like i remember writing about it just yeah probably about two or three years ago at this point and it still felt like there was a certain group of people in crypto for who it was so obvious that it was almost a boring thing to write and then for a large swath of the world like Still very, very strange and hard to wrap your head around. Right. Jamie Dimon said very famously, crypto is rat poison. And now JP Morgan has an internal blockchain where they move on to $10 billion a day on it for international settlement. So we've come a long way. And then I think the third trend that we're paying a lot of attention to is the tokenization of stock. So Robinhood is doing this and Coinbase is doing this. And I think there's two really interesting axes here. Typically, when you buy a stock, you buy it on the New York Stock Exchange or the NASDAQ. That's called a venue. And if you wanted to buy a token, let's say you wanted to buy Solana or Ethereum, you would buy it in another venue. You might buy it on Coinbase, which is a centralized exchange. You might buy it on a decentralized exchange like Uniswap. But you can't buy stocks using crypto exchanges and you can't buy tokens using stock exchanges. But the brokerage is refusing them.

So Robinhood today, you can buy crypto and you can buy stocks. Where the venue is, where the asset underlying, it doesn't matter. And then now what you can do is you can buy synthetics. Robinhood is offering shares of OpenAI that you can buy. And so it's on a blockchain. It's a tokenized stock. There's a lot to unpack there. Yes. But what does this mean? Well, it means that... What's really exciting for crypto is you have the entire US stock market, which is the largest stock market in the world. And then you have crypto, which is large, but not as large. And all of a sudden, those pools of capital should fuse, which is great. The other thing it means is that the startups should have a much faster path. I think software startups will have a much faster path to IPO as a result of crypto than they've had in the last 10 years. Oh, wow. Interesting. I want to unpack both of those. quickly, but the first part, you say they should fuse. Can you explain to folks why that almost is an inevitability that these things are going to come together from an architectural advantages standpoint? Yeah. There's no real difference between a token and a stock. What do you have with the stock? You have a voting right, depending. You have a share in the company. You have A dividend, right? Let's just say it's those three and the ability to trade it, right? So you have those three things. Well, a token is the same thing. I have a dividend, right? It can produce yield for me. I can vote and I can trade it. And so, okay, securities law, I think will ultimately catch up and say both of these things are the same thing. So legally, and there are already significant volumes of debt, like bonds, that are on crypto exchanges that are traded just as they would be on bond markets. So the bond market was the first one to normalize relations, let's say. And the equities market is coming. The other thing that's happening is publicly traded software companies trade on multiples. We talk about forward multiples, EV to NTM. And for a long time, crypto tokens did not. I mean, you looked at some of these multiples, they were in the tens of thousands of times of revenue. Yes. And over the last six months.

certain categories of crypto tokens are now trading at 40 to 60 times REVs. So still elevated, but we're now achieving normalization. And so if I'm a software investor and I have a dollar to invest, and it's just as easy for me to invest in the equity market as a token market, and I have a reasonably good proxy of financial statements between the two, now all of a sudden my dollars can flow freely and I'm satisfying my fiduciary responsibility. to do diligence on both of them, which is lacking today in the token world. There's no gap. There's no S1, right? But it's coming. How does this change end up accelerating the timeline to IPOs for some of these software companies, do you think? How much does it cost to go public? I'm not talking about the dilution or the amount of money raised. I'm talking about the legal fees, investor relations, the regulatory fees. What do you think it costs to take a software company public? Gosh, I honestly have no idea. I would put it in those sort of like single digit millions. 15 to 25 million. Okay, wow. I'm off by a good chunk there. Okay, so if you're a $100 million revenue company, is it rational to raise a round of financing that costs you 15 to 25 million? Yeah, totally. Wow. So to put it into perspective, if you were going to raise a series X, very late stage round, your legal fees might be in the hundreds of thousands of dollars. You might touch a million depending on how big your cap table is. Series A is 30K. Series B is 40K. Series C is 50K. So all of a sudden, you have this step up of like three or four orders of magnitude in just your transaction cost to go public, which is why it used to be in the late 90s, you needed like 25 million of trailing revenue to go public. And today, no, no, no. You can't go public unless you have 250, maybe 500 million in revenue because as an expense, you'd be a banana. Just raise it in the private markets. And so that's why you look at the average growth of publicly traded companies, it's asymptoting to 10%, basically, which is software inflation, because it's so expensive to go public. So why is it so expensive? Well, there's just a lot of regulation, right? Sarbanes, Oxley, all that kind of stuff. And so if you look at crypto exchanges, the regulation there is much less. There's no S1, there's no audit.

Some of those components are important, but the point is we have made it so expensive to go public that if you're a, call it micro cap software company, if you were somewhere between one and a half to 2 billion, it's not economic to go public. Most of those companies get taken out by PE right away. Wouldn't it be amazing if you had a venue where small cap, high growth software companies were publicly traded? And so that's what gets me so excited. And these exist. You look at Chinex in China or the ASX in the UK or TSX or the rumors of the Texas Stock Exchange, which would be a new venue for trading. All of those venues are focused on high growth, small cap technology companies where the disclosure requirements are significantly less than the New York Stock Exchange or the NASDAQ. Wow, how interesting. You also mentioned the Robinhood. open AI example, which is more like, you know, secondaries trading, which feels like an exciting, you know, other set of other asset class that sort of needs to be opened up. But there's all these sort of complexities about like, are these essentially options? Is it, you know, owned through an SPV and something like who owns it? How do you see that ending, like ending up playing out? Well, I've been reading online trying to understand because Robinhood mentioned the open AI tokenized stock and open AI said we did not authorize this. So in most secondary transactions, the board must authorize those transactions. And so my understanding, which is only based on what I read online, is that Robinhood is creating a synthetic asset that tracks the value of OpenAI stock, and they're backstopping it. And maybe there's an SPV underneath. Anyway, it's not clear. Unclear, yeah. We'll see how it all shakes out, I suppose. But I do think the notion of tokenized secondaries, or secondaries that float, or even tokenized venture firms where LPs can trade underlying positions is important. Let me reframe venture capital the way that PE works today. I started a seed fund. I find a company, two people, and a dog, and I help them get to a series B. And I work with them for four years, and then I sell the entirety of my position. And then a series B specialist fund that is focused on scaling go-to-market,

buys that entire position and grows the company from a million in ARR to 25 million in ARR and gets the benefit of both the revenue growth and the multiple expansion associated with the revenue growth, and then sells it to a Series E investor. And I'm just making things up. And that Series E investor is really good at preparing companies to go public. They have the right Rolodex of executives, they know the investment bankers, and they go from Series E to IPO. That's the way that a large part of the private equity market works. And there are a couple of benefits. The first is, The time for investors to get their money back, the DPI dollars to pay it in, is much faster, which makes subsequent fundraising much faster. There's specialization on individual stages. And so the cycle of money basically just moves a whole lot faster. We're nowhere close to that, right? We talked about venture capital is a $250 billion asset class. In the last three years, the totality of all secondary funds is $6 billion. So we're talking maybe a basis point. We're not there yet. But I wonder, one of the questions in my mind is, I wonder if we get to a place where seed funds are liquidating positions after three or four years. Series A funds are liquidating, maybe not the entirety, but some fraction thereof. Just $22 a month or $220 annually. So what's included? One, tactical interviews where elite founders and investors reveal their actual strategies and decision frameworks. Two, comprehensive guides that distill hundreds of hours of research into actionable insights on investing and company building. Three, an exclusive database of emerging startups poised for significant growth. And finally, complete access. to our archive of meticulously crafted case studies. All of this comes wrapped in the distinctive storytelling and incisive analysis that readers have come to expect from the generalist. We've designed Generalist Plus to level up your capabilities as an investor and operator through knowledge that matters delivered with precision and depth. So join a community of strategic thinkers who are gaining an edge in understanding markets, technology, and business fundamentals.

by visiting thegeneralist.substack.com. That's thegeneralist.substack.com. You know, we were talking about some of these crypto innovations and, you know, I'm reminded of a piece I think you wrote in early 2024 where you talked about the most profitable software startup in the world is Ethereum. And I thought that was such a good framing and such an interesting way that you talked about it. It feels like it's been a really sort of tough cycle for Ethereum in general. There's lots of soul searching, lots of discussion about where things head. How have you sort of looked at that? How do you see its place in the ecosystem? You know, I do think there's such a clear story around Bitcoin, as you said, and a clearer one around Solana. But, you know, are people... two down on Ethereum? How do you think about its potential today? Yeah, you had a brilliant point in there. You said Ethereum soul searching. So just for the audience, Solana's nickname is Sol. There you go. That's right. And so the major competitor to Ethereum is Solana. So I think Ethereum is kind of the grandparent of all of these chains. And the volumes, I mean, just to kind of put it in perspective, 50% to 60% of all stablecoin volume is on Ethereum, and another big chunk is on Tron. So it's really the vast majority part of the market. And then 60% plus of the decentralized finance or DeFi market is on Ethereum. So it's the big dog. And Ethereum, it plateaued, right? There was not a lot of innovation. They allowed other companies to innovate on top, which are called layer twos, like Arbitrum and Optimism. And they were solving, those L2s were solving performance challenges for the ecosystem. And so they were basically allowing you to move money faster and cheaper than Ethereum and saving you 90% to kind of give you a sense of the relative efficiencies of those systems. And then Solana came around and said, we're going to architect in a very different way. We'll be more centralized. It will be faster and cheaper. And then you have another one called SWE, which is lucky enough to be an investor that also has very, very significant performance improvements as the X meta blockchain.

team, one of the two. And so within the span of three years, you had two orders of magnitude improvement in blockchain performance relative to Ethereum. Let me ground that in an example. If I wanted to send you a crypto kitty, it would cost me $50 to $70 to send it to you on Ethereum. If I used an L2 like Arbitrum, it might cost me $2.50. And then if I use a more modern blockchain like Solana or Sui, it might cost me 20 cents, right? So very significant reduction in transaction costs. And so as a result, over time, well, developers will build on the systems that are more efficient, much better price performance characteristics. Nobody wants to pick those fees. And so that's where we found ourselves with ETH maybe three to four months ago. And then people started all of a sudden- Waking up. Yeah, dissatisfaction. The community decided- okay, we need to react. So they raised $150 million and Vitalik is more involved and they're at this crux, this critical moment in the blockchain's development where they need to figure out what their competitive advantage is. It's a little bit like AWS before AI, it was the standard. Everybody, I mean, it was not even a question if you were a startup, which platform you're building on AWS. And then AI came around. No, Azure became the dominant platform. Why? Because they had a special relationship with OpenAI. And so Solana has a special relationship with Stripe, right? Robinhood has a special relationship with Hyperliquid. And now all of a sudden you have bigger strategic relationships that change the builder's perspectives on which platform to build. And so Ethereum finds itself where AWS is or was six or seven months ago. There's been this huge wave of innovation, other people have pushed, and now the developer mindset is shifting. As you think about its comparative advantages and what should be the real advantages of the project, what do you think that should look like? What do you see as the things that the grandfather still has or should lean into that these younger chains?

maybe can't compete with or will struggle to compete with. The single most important thing in any financial market is liquidity. And Ethereum has it now, right? So the question is, how do you capture that liquidity? Well, stable coins will be an absolutely essential market. And minimizing the transaction costs so that liquidity remains on ETH is priority one. How do they do that? What do they do with the L2s? How do they think about the consensus infrastructure? And then the reward systems for the validators? I think is a really, basically, how do you architect your entire system? Yeah. They'll have to, I think, revisit a lot of the core assumptions because the underlying technology has changed. And that's okay. But I think a lot of times in the world of Web3, we have attachments to certain ideas. And over time, they change. Really interesting. As you sort of look at the... Crypto Web3 landscape, are there newer projects that you've spent a lot of time playing with that you think are really promising or new sort of pieces of infrastructure that you think are worth discussing that aren't being as discussed right now? Well, I mean, there's a brand new company called Hyperliquid that's worth somewhere between 20 to 40 billion. That outside the world of crypto, it's not very well known, not venture backed. And they did a couple of things that were absolutely brilliant. One, they focused on the perpetual market, highest margin product. They created a blockchain that has some unique attributes to that market in particular. So they specialized a vertically integrated stack exclusively on it. And so as a result of the vertical integration and then making the blockchain specific and the tokenomics and all that stuff specific to what they were doing, they built a huge, I mean, multi-decacorn business that's now partnering with Phantom, right? So Phantom now you can trade perps and it's all powered through Hyperliquid. And I think turned, as you said, no venture capital and turned down a lot of venture capital as far as I know, which is pretty cool. Really cool. $30 billion bootstrap. Yeah, pretty amazing. And then they're doing all kinds of other things where they buy their tokens back. It's like a stock buyback program, just like Apple. And so they're actively managing the value of their tokens. There's a lot of foresight and they deserve a lot of credit for the company that they've built. So I think the vertical integration.

is one really interesting trend. I think another big question is, are we moving from proof of work to proof of stake to proof of authority? Can you explain those for folks just very briefly? I know that's not the easiest, but in simple terms. Yeah. So Bitcoin is proof of work. If you want to earn money by minting Bitcoin, you need a really big computer and you basically solve very complex, not complex, but computationally intensive math problems. And then you're rewarded with a fraction of a Bitcoin. That's proof of work. Proof of stake is someone is vouching for me. If I'm trying to join a, I don't know, whatever, social club, right? And Mario vouches for me, that's proof of stake. He's staking his reputation. People do that in the financial world. It's like, I know Mario's trading. I've got him. Yeah. That's what it is. Backstopping you in a way. Right. And then proof of authority, and I'm rewarded for that, for backing up, Mario. And then proof of authority is... I pick Mario. I think Mario is a great guy. That's enough. Mario will run his computers. He will validate the transactions and I trust him, which is how most database today's work, right? If I send data to Amazon, I trust Amazon will run the database the right way. They're not going to change how I move money or in my Quicken, however many, they won't change it. I just trust Amazon. So proof of authority with one. validator is what we have today. And so we're kind of moving to this world where we have less and less decentralization. But ending up with more efficiency as the sort of benefit of it. You trade off decentralization for speed, efficiency, et cetera. At some point, you get to the place where there's only one validator and you basically have replicated an existing database. We were chatting with Patrick O'Grady. yesterday. He was one of the engineers from Coinbase about it. And so the question is, well, why do you have decentralization if you only have delegated authority? Well, you might have delegated authority. We talked about JP Morgan. Within the world of JP Morgan, they choose who validates the transactions and says this transaction is valid. So they're already running it internally. There are a set of applications where this might work. So Ethereum has to figure out where on that spectrum of consensus they want to be. And I think this is a...

It's like another kind of critical moment for the ecosystem to decide what kind of consensus mechanisms. Amazing. Well, I want to talk about AI a little bit. We've talked a little bit about it, but in reading your writing, it's so clear to me that you're playing with these technologies so much. You're building stuff with AI. You're using voice so much. In your own sort of day-to-day life, what do you find are the most surprising places or the most surprising use cases that you do rely on these technologies for? Okay, so I spent the entirety of the July 4th weekend trying to only interact with my computer through AI. So there's a tool called Claude Code. Gemini has another one where you can just type in and say, do this and do this and do this and do this. And I'll give you an example. So you can just ask it a question like you would on Claude.com, Claude.ai, whatever the domain is. And you can ask it, you know, what is the difference between, this is a query from last night, SEAL Team 6 and Delta Force. I had no idea. And it will answer you. But then you can also tell it, hey, look through all my files and see if you can find this blog post from two years ago. Oh, wow. It's another very basic thing you can do. But then you can ask it, create a tool. So let's create a tool. What is my schedule for tomorrow? OK, so I asked it just to write the code itself. And then it will go into my Google Calendar. And then it'll spit it out. And then I created another tool which said, go and get my email and summarize it for me. And then I created another tool which said, tie this to Asana, a task manager. And so before I got on the podcast today, I went through all my email at once. And I looked at all the email and I said, do this with this email and create a task from this one and add this one to the CRM and check to see if I can meet this person at this time. And if not, send them the slots. And I can do that all with one voice command. And so what I'm trying to understand is what are the limits?

Right. If we were going to reimagine a computer that I was only speaking to that had access to all these different tools, where does it break down? It's so interesting because in some ways you've ended up going back to the start of computing in some way, like you're back to almost a command line just with voice and one tab that's doing all these things for you. What are the limits that you've discovered of it? There are a couple of limits. The first is I really want the computer to go off and do something and then come back when it's done. That does, it's not very well handled today. You can use some more technical tools to do that. The second thing is it often makes mistakes. So I've given it a tool to send email. Well, sometimes it just decides to send an email, right? Or sometimes it just decides to archive an email. And so you have to instruct it in its memory to say, never send an email or archive an email without my permission. And so there's this, it's like training. a brand new person. How do you work? So there is a lot of constant feedback. And the bet I'm making, and right now I'm 50-50, is that within a week or two, all of that training will result in improved productivity. But I couldn't tell you that it's doing that today. Wow, but it's great if you only have a couple weeks to figure it out. You think by then you'll have a reasonably definitive answer. Because then the tools exist and then I find all of the different cases where the tools fail, right? So like yesterday I was asking it, what is my calendar for tomorrow? It turns out it was only fetching the first 10 calendar items. And so it was making up nonsense, right? And so anyway, there are all these corner cases. So maybe over time. And then I wonder if companies themselves will have these shared libraries where everyone is accessing the calendar through some, and this is actually what we're building inside of theory. Adam, our great software engineer, he released the library for the CRM yesterday that's available to AI. Okay, this is exactly where I wanted to go next, which is how are you using AI within the firm? Because I know that you guys are a very technical team and thinking about how to architect the firm in a way that involves a decent amount of engineering. The most important thing is that we understand these technologies at a very deep level and feel the same problems that builders feel.

Because if we do that, I think we'll be better investors. The second thing is there's a lot of work that can be automated. And large language models are really great at this kind of unstructured automation. So I mean, think about it just like a sales process. You have a lead. You want to process that lead and qualify that lead as quickly as possible. Go and find as much information about as possible. Synthesize. We have an agent that builds market sizing analysis. We have an alpha of an agent that makes investment recommendations that I wouldn't trust. You can kind of, you can, we're starting to assemble all of those pieces, memo generation, deck extraction, those kinds of things. And reality is none of these technologies are flawless and they're unbelievable improvements in productivity. As much as we look at startups and say, okay, the ARR per engineer should now go from about 150K to 500K. Well, the number of startups. covered by a VC investor should go up by a similar percentage. There's no reason it shouldn't. So maybe we can cover three times as many spaces or four times as many spaces, and we should be able to do it with greater depth. What is the entire toolkit that's required to be able to do that? I enjoy creating cloud projects for investments I'm considering and giving it inputs to say, try and analyze whether I should make a decision or not. I find it's really not very good at at least mapping what my own judgment would be. But it's really interesting. And still, I find it quite productive almost as a conversation to have with an intelligence of some kind. What do you find that's like for your sort of investment recommendation engine that you have or your agent? What does it help you with? Yeah, I think summarization is really good. In a space where we don't know a lot, let's say all of a sudden we decide to invest in chips. It helps you get up to speed very fast, much faster than you could normally. Sometimes it surfaces questions or insights from history that are important. It's really a summarization that it's quite good. Ontology generation, it's really good at. So let's say you were looking at MCP servers and you wanted to classify all the different approaches to different model context protocols, next generation API servers, basically.

classify them and segment them. Or you wanted to look at an even broader universe, AI security companies. Well, there's data loss prevention companies, firewalls, prompt injection company. Anyway, so it's also really good at that. Well, one of the last times that we chatted, I remember you saying that one theory has like a very concentrated approach where I think you were sort of aiming for 12 portfolio companies per vintage. And two, that the way you'd sort of come to that number was by running like Monte Carlo simulations to figure out what is the optimal size. What are the other ways you rely on data or this type of simulation to make sort of like key strategic decisions around the contours of the firm? So yes, that's right. We did that. And then Carter released an analysis, I'm not sure if you saw, of the... The return distribution of funds is a function of their portfolio construction. It's really interesting where it shows concentrated funds, median is less, but they're 75th to 90th percentile or higher. So it's a higher vault, different kind of return profile. I think within the world of early stage, it's really hard because there's not a ton of data. There's not a lot of structured data. There is a lot of unstructured data. So what people say online about different products or spaces is fascinating. And AI is really good at capturing that. What people say in podcasts or recording our interviews or conversations that we have with people, that's where the majority of the data is or the majority of the insight. How do you ask the right questions? What kinds of interviews? Which are the sources that you pick? So I think that's where we use it. And then the other place that we use it is benchmarks and comparables and those kinds of places. It's not like hedge funds, right? We're not buying satellite images over retailers to predict quarterly earnings. In your sort of introductory post about Theory Ventures, you talked about investing in machine learning as a force multiplier, which feels like the lens through which you look at AI in general. Why was that the way that you...

perceived the AI opportunity and the right way for theory to play it exactly. Let me ask you a question. Okay. So on a manufacturing assembly line for cars, how many humans does a robot replace? I would guess two, one, I don't know. Two and a half. Two and a half? Yeah. Okay. In an Amazon distribution center. So that was my mental model for AI. AI would kind of replace two and a half people inside of a company. Amazon then released the stat, this is three or four months ago, where they said in their distribution centers, their warehouses, one robot replaces 20 humans. Wow. Gosh. Why is a robot in a DC that much more productive than a robot on an assembly line? Because a robot on an assembly line, it's the same thing. You're welding two joints or attaching a windshield, whatever it is. And so I don't have the answer to it. The broader point though is, okay, well, if you're a knowledge worker, if you're a white collar worker, what is that ratio? Is it two? Is it 20? Is it 100? Is it 10? And so the question behind the question is, how many agents should a highly productive worker manage? I think right now I can manage two to four. And the highest number I've heard is 15. I bet within the next 12 months. We will see workers managing 100 agents working simultaneously. What does the person who manages 15, how do they do that to the extent that you can share? What does that look like? They have a task list. It's a software engineer who creates a task, a PR, a pull review. It says, go and fix this. Go and create this feature. Go and fix this. 15 times. Comes back half an hour and then reviews them. There's no reason that that mental model of working shouldn't apply to a salesperson. Or it shouldn't apply to a marketer or a customer support rep. And so, I don't know. I think the question in my mind is, and this is the reason why I'm playing around with AI on the computer, is how do I get to a place where I can manage 50 times more agents than I can today? And right now, you still need to monitor and watch. And you're watching the readouts and making sure that it's working okay. Yeah.

There has to be a system. There has to be one, an inbox, which Harrison from Langchain talked about, an agentic inbox to manage all these things outside the world of software engineering. And then the second is there has to be some kind of way of putting guardrails around them so they don't send 100 emails to the wrong people on your behalf. Yeah, it feels like you need new systems to manage it on the human level and also just like a different level of reliability such that you can really trust it to do what you ask it to do. Right, exactly right. About a year ago, you wrote that Microsoft was leading the AI race and Google was lagging. How do you think that has changed in the 18 months since? Well, a lot's changed, right? So the relationship between Microsoft and OpenAI is now quite different. The Gemini models are all state of the art. You look at the efficient frontier of performance as a function of model size, and the Gemini models are all either number one or number two. So they're meaningfully I mean, it made a huge leap. I think Google is clearly subsidizing a lot of their AI products to drive usage. So they are maybe in the developer world a bit behind, and they're trying to drive more adoption. But I think to their strength, look at the impact on their search business. It must be phenomenal, an improvement in margin, because there are two different kinds of ads that Google runs. One is the ones on search and the other ones are on other people's webpages. And the traffic to other people's webpages is plummeting, right? So HubSpot's traffic's down 70%, 75%. And other sites are seeing more and more declines because of those AI overviews. Well, great for Google because now all of a sudden there are more and more ads where they're capturing 100% of the ad spend as opposed to some fraction of it. So that's good. So I think on the consumer side, they're in an unbelievable position, an absolutely enviable position. And then the other dynamic that a lot of people don't talk about is access to the GPU, right? Running these models is super expensive. Chrome, a browser, the dominant browser can actually use a GPU. So they can put a model in your browser and then have that use your GPU to answer queries and they don't have to spin up anything in the data center. A lot of the other model companies do not.

have that advantage. Yes. So I think on the consumer side, they're in a great position. And the enterprise side, they're still, they have unbelievable products and they're solving the distribution, which is the classic Google strategic challenge. Which are the models that you end up using most often? I use a lot of open source models. I use a lot of the Google models because they're free, but it's different for different things. So Gemini is my default model. OpenAI, I think. is great at the deep research product I really like. Yeah, agreed. Yeah, it's really good. And then Claude, I used to use a lot for writing. And now when Gemini cannot solve a coding problem, Claude fixes it every time. Oh, wow. Interesting. How do you think about OpenAI's position today? And I'm especially sort of thinking about it with respect to the aggressive moves that Zuckerberg's been making and especially on the hiring front. To that end, where do you see meta in this? They've obviously taken a very different approach. Yeah, great question. I mean, OpenAI has an unbelievable asset, which is hundreds of millions of daily active users and a brand that is universal, right? And that, I think, is an asset that is hard to assail. It's hard to attack. So I think they're in a relatively strong position because of that consumer distribution. Any company that reaches that level of scale is worth at least tens of billions, if not hundreds of billions, if not trillions. Google got to that level of scale. Meta got to that level of scale just on consumer distribution. Snapchat got to some level of scale on that distribution. All those companies are worth a lot. There's a long-term business model question there that they're working to solve. I think their pricing power is... underappreciated, look at cursor increasing their pricing. I think a lot of people, especially in the business world, will pay a lot of money for these tools because of the productivity gains. And so there's a pricing discovery opportunity that is there. But they used to have a trillion dollar market cap company in their corner and now they don't, right? And they have Masa instead. And so they're figuring it out, right? They need to figure out who their allies are.

what the long-term positioning is, but ultimately the brand and the distribution is unbelievably valuable. And by comparison, how do you think about these aggressive meta hiring moves and how it sort of positions itself? Because it's had to play a bit of catch up and is doing more of the open source stuff, but obviously has so much of that distribution it can rely on. Okay. So the first thing I'd say is how awesome is it that you have the CEO of a $2 trillion company being that aggressive. I know you do. I think it's the most, you do not see people play offense that hard like that I can ever think of. Yeah. I mean, just overt frontal assault, right? And I'm coming. Yeah. You have to respect that. That is awesome because you know, and then I was reading the announcement today with Zuckerberg where he invested three and a half billion into the largest eyeglass maker, Luxottica. You know, he bought three to 5% of the company because that's a new And so I just love the aggression. I think it makes business fun to see people conduct business that way. Yeah. Go that hard. Yeah. I think Lama were the bleeding edge of open source models for a long time until Lama 4. I can't tell what happened there. Yeah. But I used to use a lot of open source Lama models and now I don't anymore. It's a Gemma model. And so I think there's some catch up to do there. But given the amount of talent influx into meta, I think give them six to 12 months and I think we'll see some pretty spectacular results. Amazing. Well, we always like to sort of wrap up episodes with a few more philosophical questions. So I think these will be fun. If you had unlimited resources and no operational constraints, what experiment would you like to run? I would build a computer that was entirely AI. Like a PC? Yeah. The only thing you could do is talk to it. I think that'd be a lot of fun. That would be a lot of fun. Would you have a screen at all for output? I think you need a screen sometimes, but not all the time. So I would optimize it for voice. And then when you need to see an image, maybe you have like her, right? Like he pulls out a little screen from his pocket to look at a photo, but then he puts it back and it's all voice. And I think, I remember meeting this brilliant founder from MIT who was designing, he was designing a

human computer interaction system for disabled people using their tongue. So he gave you a retainer that would go in the top and then you would manipulate the mouse by using the tip of your tongue against it. And I think we need to reinvent. That was just an inspiration for a completely new form of human computer interaction. I think there's an opportunity to completely transform human computer interaction again. And so I think it'd be a lot of fun to embark on that journey and figure out what breaks where the sharp edge is. Yeah, I 100% agree. Have you played with any of the sort of I don't know, new AI hardware consumer-y products like, you know, I'm thinking Rabbit or Humane, you know, back in the day and stuff like that. Yeah. I mean, they're all steps forward, right? It's like the magic link before the iPad. Yeah. I think glasses are probably the new form factor. And I'm really excited about the Meta Orion glasses, the little projector and the fact that they listen all the time and the AI is there and it doesn't look like your phone. But the reason I brought up the story about the MIT... founder is, he says, I want to make sure that people spend just as much time with computers, but less time on screens. Yeah, that's also, you know, using the sort of tongue movements. You can even imagine that for people who don't have disabilities, like, you know, there will need to be some sort of sub vocal version of these discussions if these become our dominant computers, right? You know, whispering or whatever it is. Another friend of mine said, he had this really great. prompt, it was Nikunj, sorry, Nikunj Katari said, do you believe you're biotic? And I said, no. And he said, well, when was the last time you forgot your phone anywhere? You already have a computer attached to you. Yeah, totally. So we're close. Yeah. Yeah, absolutely. What's a tradition or practice from either another culture or time period that you think we should more widely adopt? I was listening to a YouTube video of a Shaolin master. And it's been banging around in my head since then. He said, When you eat, you eat. When you work, you work. When you sleep, you sleep. And so that single focus, I think it's harder and harder, but I think it's really important. So I would love to see us do more of that, but it's hard. Yeah, that level of presence is something to aspire to for sure. Okay, sort of last question. If you had the power to assign a book to everyone on earth to read and understand, what book would you choose? One of the most formational books in my life.

is this book called Narcissus and Goldman. And it's written by a German guy or Austrian guy named Hermann Hess. And he was an European who brought Eastern philosophy to the West. And the book Narcissus and Goldman is a story of two boys who meet at a monastery in the Middle Ages. And they go through very different lives. And then they meet on the deathbed of one of them. And they reflect. And I don't... There was an amazing lesson in there. I read it. My aunt gave it to me when I was 14 or 15. I have never forgotten that book. And there's an amazing lesson in there, which is there are a thousand different ways to live your life. And at the end, you kind of meet the people, you meet again and you realize you kind of get to the same place. So it's all about choosing that path. Just be very deliberate about the path that you choose. And so that was just because one of them is an ascetic. He abstains from everything. And the other is a bon vivant. He lives life to the max. And they end up in the same place. I mean, for that moment in time, but I would highly recommend that book. I'd never heard of it. I've read some of Hermann Hesse's other books, but not that one. So that's an awesome recommendation. Thank you so much for doing this, Tomas. It's been amazing to chat. And yeah, I'm so grateful for you sharing all of your knowledge with me. My pleasure was mine, Mario. Thanks for having me on the show. That's it. Thank you for listening to this episode of The Generalist Podcast. Please subscribe on Apple Podcasts, Spotify, or your preferred podcast app. Ratings and reviews help others discover these discussions. So if you enjoyed the conversation, I'd be grateful if you could take a moment to leave one. For all past episodes and more, visit us at thegeneralist.substack.com. See you next time as we continue to explore the future.