Ben talks with venture capitalist Tomaz Tunguz about his path from Java engineer to Google product manager and finally to author and investor. They cover why lower data storage costs and improved database performance are driving a new wave of innovation, how LLMs learn by doing, and how companies (and investors) should respond to uncertain economic conditions.
Tomasz is a general partner at Theory Ventures, a venture capital firm focused on early-stage software companies.
He coauthored the book Winning with Data, a deep dive into how big data has changed business best practices and organizational culture.
Find Tomasz’s writing here.
Follow Tomasz on LinkedIn or Twitter.com.
In honor of Tomasz’s early career, we’re shouting out Johnny Hujol’s answer to What exactly is a container in J2EE and how does it help?.
[intro music plays]
Ben Popper 70% of the Fortune 500 use Pluralsight to upskill their workforce. Now you can take the same courses to boost your dev skills. Start a free trial today. Visit pluralsight.com/stack to learn more.
BP Hello, everybody. Welcome back to the Stack Overflow Podcast, a place to talk all things software and technology. A lot of what we build in this industry and what software developers spend time working on begins as a startup, and the money for those startups more often than not, unless they're bootstrapped, which is great, comes from venture capital. We are lucky today to have on the program Tomasz Tunguz, who has experience both as an engineer builder and also as an investor. So we're going to chat a little bit about what he's seeing in the market, what's going on with AI, and making that jump from practitioner to investor. So Tomasz, welcome to the program.
Tom Tunguz Pleasure to be here. Thanks for inviting me.
BP So for folks who don't know, give them a little background. In the past, you worked at places like Google as a product manager, you were an engineer before then, and then you kind of transitioned from there into becoming an author and an investor. So maybe tell folks a little bit about that journey.
TT Yeah, you bet. So in college and grad school I studied mechanical engineering and computer science and really fell in love with control systems. And then my first gig out of grad school was a Java engineer working for the Department of Homeland Security and we were building different internal systems there and that's what really got me into the world of coding, using Stack Overflow a ton in order to understand J2EE and all the different frameworks. It was Spring and all the ORMs, Hibernate and all that stuff. And then went to Google and was a product manager, started as a customer support rep, built two different internal products. That's when I learned to use Ruby and fell in love with that stack in 2005-2006– Ruby on Rails. And then I managed a team that built large scale machine learning models for advertising targeting in many different languages, but I wasn't coding them. And I was at Redpoint for about 14 years and invested in many, many different companies. I still code every day, take my notes in markdown and build lots of different internal tools in Ruby and Go, because Go is actually super high performance, although we can talk about the error handling another time, a lot to be desired there. And then started Theory, and so Theory is a $235 million first time fund. We have six people on the team, all of whom are focused on data with technical backgrounds. And we spent a lot of time in software applications, machine learning, and we're playing around with large language models. Two weeks ago we employed Llama 2 and hooked it up to Whisper so I could start talking to my computer and have it start doing things, and ran into lots of memory issues.
BP Yeah, I'm excited to talk to my coffee maker. That's the future I've been hoping for. So folks know you mentioned you started Theory and that's a relatively young fund. You said first time fund, but I have to assume based on your LinkedIn that you were investing before then. You've had a lot of pretty notable names on here where you were a board member that went to an exit along with some acquisitions. I think people who are listening probably heard of Expensify, probably heard of Looker, and I could go on and on. So before that, were you at a different venture shop?
TT Yeah, I was at Redpoint and worked with eight unicorns there in a bunch of different spaces, a couple at the application layer, Expensify in customer a lot and the data stack, we worked with Looker, MotherDuck, and Hex, and then two in crypto and Web3. So I look at Web 3 as there are many innovations there, stablecoins and DeFi, but the one that gets me most excited is decentralized databases. I look at Ethereum and Ethereum is worth about five snowflakes put together. It's really big and probably the fastest growing database company of the last 20 years. But I don't think people look at it as a database company, they look at it more as a token.
BP I'll put you on the spot and let's defend that a little. I think I started reporting on technology in 2010, so I missed my chance to retire for life. I think we spent our Bitcoin when it got to $20 each. We were like, “This is the top.” We went out and bought some pizza at Charlie Shrem's place in New York, but that's okay, I have no regrets. And then we went through a crypto boom in the 10 years following that and an amazing IPO for Coinbase and something like a $3 trillion valuation across that market. And then we've had sort of a pretty harsh correction and are going through the biggest fraud case in US history with FTX at the moment. So when you think about Ethereum and Web3, ironman for me the point that this is not about cryptocurrency, this is not about tokens and NFTs, this is actually a database play in some way.
TT Yeah, that's the way we view it. I think these databases work in a different way where you don't have to trust anybody and what they're enabling is the fastest way of moving money across borders today. And the other things that they're enabling, we've met companies that have optimized stream processing using these decentralized technologies, whether it's consensus protocols or multi-party computation, they're able to get significant benefits from parallelization. And so it's not that they don't trust the core central actor, but they're able to take a bunch of these technologies and actually create a new architecture. And this is sort of very high level, but the way that we've been talking about it internally is distributed and decentralized systems converge over time, and a lot of the innovations for those decentralized systems will become part of distributed systems.
BP I have to agree on the first point. We've had some folks come on the show and talk about cross-border transactions and seem to have built a good business around remittance. And then obviously if you live in an authoritarian state and you want to get your money out, you're glad that crypto exists. On the second one, when you say it has benefits for– did you say streaming and converging on a truth, I have heard that discussion in the realm of, for example, video games. You've got this first person shooter and there's a bigtime tournament and they're playing on different sides of the globe and they need to resolve whether or not that was a hit or miss to decide who wins the golden cup. In what way does a Web3 blockchain-based company help with that?
TT So in that use case, which is a UDP use case, I'm not as deep there. But what we have seen is, let's say you're processing a big stream of data and what you're able to do with these decentralized systems is actually move to far more significant parallelization. That's one use case. Another application is federated machine learning, so say you're a big bank and you want to ensure that you run an anti-fraud algorithm exactly the same way across five different branches of the bank in different parts of the world where there are different data locality regimes and you need to guarantee to a regulator that you did exactly what you said that you did. Then I think to have a distributed orchestrator and that distributed orchestrator can then write a ZK proof that demonstrates exactly what you said, and that ZK proof is written to a public file system and that public file system is then interrogated by a compliance. And so that's definitely one application of it. I think talking about the data locality rules, we envision a world where in the future, the cost associated with custodying data becomes so significant that many software applications that are touching large amounts of PII are re-architected where the users themselves are the ones custodying the data and providing selective access. So imagine a Salesforce that's built on top of NFTs that contains wiring information, and those NFTs are exposed only selectively, and whenever that wiring information changes, it's the person who's wiring information changes, and then that cascades through everybody else's accounts. So what is the timeframe for that kind of massive architectural change at the software level? It's hard to say. It's not 3 years, maybe it's 5, maybe it's 10, maybe it's 25, but I do think we're getting to a place where the cost of regulatory compliance will become so significant and the pendulum swing will go back to local first applications with people desiring to control their own data.
BP Right. I read a post you wrote on it discussing many of the same things I had about why we're kind of in this bubble bursting moment, and I have to agree with your pretty sanguine take here. The gloom will remain over the ecosystem until somebody figures out a way to assemble decentralized databases, tokenomics, wallets and NFTs into a product that changes the world. So we're still waiting on that. That's an unknown known, in your opinion.
TT I think AWS cut their prices something like 120 times within the first three years, and I think we're sort of at that place in the ecosystem and Web3, where it used to cost 50 bucks to store a bit of data on Ethereum and now the next generation technologies are doing it at a fraction of a penny, so the less and less it actually ultimately costs to store data and the better the performance of the underlying database opens up a universe of different applications that maybe a year or two ago were not possible because of the latency or the cost.
BP Interesting. So let's change gears for a minute. I think you said that part of your portfolio is looking at Web3 and another part of it is looking at AI. You mentioned that you're playing around with these things. If you're at Stack Overflow, you're obviously no stranger to the impact it can have and the potential as well, so tell me a little bit about your perspective on this market, what you think might be exciting for software developers who are listening, and based on that, what kind of investments you're making?
TT So as a software developer, I use a terminal based email client called Mutt, and within Mutt I use Copilot in order to kind of complete the sentences and that's been fantastic. The other places where I've really noticed the applications of LLMs in my coding is being able to learn best practices. So I learned Ruby from a bunch of books and I've never really understood the math function. I just couldn't tell you how it works. But I know it's really powerful and I read a lot about it, but it just doesn't click for me.
BP It didn't grock, yeah.
TT And then one day I was just writing code and it showed me how to do it and it was like, “Oh, that makes sense.” A third example is that I analyze a lot of data for the blog that I write at tomtunguz.com and I grew up learning R. I chose R because it had the best plotting library which I think it still does, but the reality is that the whole ecosystem has moved to Python, driven by data science and data movement, and now I need to do a lot more Python and I've been learning that language just because I'm like, “Take this R function and make it a Python function,” within a chatbot.
BP It is really amazing doing that kind of translation. You can say, “I'm used to working in this language. This is some of the code I've written. Can you show me how you would convert this over?” And as sort of a Socratic method instructor, it can be great. My kids are doing their long division right now and struggling with it and ChatGPT is a great way for them to get some screen time which always makes them excited, and then go back and forth about how this works and why this mistake can be cleared up. And it does seem like a great tutor, a great teaching aid. That seems like one of its essential functions at the moment.
TT I agree. It's infinitely patient and it will describe to you in two and three or four different ways with examples, and on the other hand there are a lot of challenges with it. One challenge is if you describe an overly complex problem, it has a very difficult time breaking it down. If it sort of makes one mistake in the sequence of steps, it has a very hard time coming back and correcting that. So I think you really have to learn how to use it. It's a tool. I think a lot of people view these systems and anthropomorphize them as having true intelligence.
BP It is not an all knowing sage. If you understand one shot, zero shot, chain of thought, and critic actor models, and you use all those techniques when you're working with an LLM, you get way better results. If you just go to it and say, “How are we going to solve peace in the Middle East?” and it doesn't give you a great answer and you see this thing as useless.
TT They're word prediction machines. They're very, very, very, very sophisticated word prediction machines and oftentimes they can lead to incredible results, but there's no– what's that chess game with the man underneath that everybody thought had a machine?
BP Yeah, Mechanical Turk.
TT Yeah, the original Mechanical Turk. There's still a lot of human involvement in producing the right kinds of outputs.
BP So you're not of the opinion that I'm just a word prediction machine at some level and that as these things get better these emergent abilities they have showcase to what degree language alone can constitute intelligence.
TT Somebody described it to me this way, which I thought was a good analogy. You can read about riding a bicycle and you can understand what it means to ride a bicycle– I have to pedal– and then you learn by doing. And I think with these large language models it's clear they can do the first. I think some people have described them as having these emergent behaviors, which goes back to von Neumann and really simple rule-based systems that produce complex emergent behavior. And maybe we're starting to see that, but I'm not convinced yet. There's some argument that as we get 100x the number of neurons in a bunch of these models maybe we'll approach it.
BP Right. I take your point and I think it's valid, which is that it's a brain in a box. It doesn't have a body or a sense of touch, although there's been some great releases recently from Google and other places where they laid an LLM on top of their robotics and adding that level of language capability made the robots a lot easier to interact with and it made them a lot more flexible in terms of the commands you could give and how they would act on them.
TT I think voice is the most natural interface for all of us to interact with a computer. We can speak three times faster than we can type, and I've been using dictation for the last 10 years to do blog posts and books and emails, manipulating my computer in order to switch between applications. I love it, I feel fast, feels like a great sense of connection.
BP How do you edit with dictation? That's my only complaint. How do you go back and edit over?
TT So I use a Mac and Dragon used to be the product and now it's embedded in the operating system. I'm not sure how they did that. But there's a product that's called Voice Control and you can say, “Change the word ‘Ben’ to ‘Tom’,” and it will go back and highlight and swap it out. Editing is still cumbersome, I'm not going to say it's lightning fast, but if you need to rip out an email or you're going back and forth on Slack, it's just that much faster. And for any of us who spend a lot of time on a keyboard, the RSI carpal tunnel savings, which is definitely true in my case, is definitely worth it.
BP No, I'm in the same world as you. I've had a lot of tendonitis and carpal tunnel and I've been dictating. The only thing I find is that if I'm writing something longer than an email, often blog posts, whatever, I lose a certain amount of facility fluidity if I have to edit with voice versus with my hands. It'd be fun to solve for that.
TT Totally. Yeah, dump it into an LLM.
BP Yeah, ask it what the answer is. So moving out of the theoretical into the more practical, you've written a couple of interesting blog posts recently about the impact of the macroeconomic climate. One was about reduction in R&D, and another was about what interest rates do to a company's runway. So what are you seeing out there in terms of how businesses are adjusting? We're recording this podcast in the context of Stack Overflow announcing we're going to do a ref, focus on certain products and try to get back to being profitable which wasn't our focus because that wasn't what the market was asking for. They were asking to see growth, growth, growth and so that's what we did and now we've got to turn in a different direction because the market has as well. So talk about the macroeconomic picture and a little bit about what that means from your perspective as an investor, and then for the folks who are listening who might be startup founders or entrepreneurs at a company, or they might be independent contributors or engineering managers wondering how their company is going to get through this, how would you look at it?
TT Yeah. So the US printed about a third more dollars than have ever existed over the last 10 years. Just to put it into perspective, we printed a lot of money and when you print a lot of money, each dollar is worth less. And then all of a sudden what we needed to do is prevent inflation from happening so we raised the cost of capital. So in order to raise a dollar of equity financing or borrow a dollar in debt, it costs you more than it did over the last 10 years. And the net result of that is companies are worth less than they were and they need to be more efficient than they have been. And there are these very painful consequences like the one that you're talking about at Stack Overflow where all of us are trying to do more with less and that means that the growth rates that venture capitalists expect from companies is changing. So when I started in the business in ‘08, going from 1 to 2 or 1 to 2.5 was a really good growth rate. And in 2018-2019 we started to see our top decile companies were growing 1 to 4, 1 to 5, every once in a while you’d come across a 1 to 7. With some of these LLM companies you see growth rates much higher than that because of the tremendous concentrated interest in a particular moment in time. And now we're at a place where, like you said, it's not growth at all costs anymore, it's how efficient is this growth? How sustainable is it? How long can you stretch your runway? To put it in burn perspectives, in 2008 a high burn was 250K a month. And there were some companies in 2018-2019 burning 3 to 5 million a month, which you can't do that anymore. Maybe if you're an LLM company and you raise a couple hundred million bucks it's possible, but most companies can't sustain those burn rates. And so those were all really big changes where unit economics matter more probably at the A, where they used to matter kind of at the C or D or as a company was approaching an IPO. That has a bunch of downstream effects.
BP Does it make being a venture capitalist less fun? It used to just be funny money. The valuations always went up, the funds always got raised. Now it's tougher.
TT Everybody's happy when we're all making money. Exits are easy, up rounds come fast and furious, and the value of a company is divorced from any sort of economic value. It's more like a rare asset syndrome.
BP Popularity contest of mindshare.
TT Exactly. So I think what it does for the entire ecosystem is it demands a different level of discipline or integrity or honesty about the kinds of businesses that we're building. And the way that we work at Theory is we spend a lot of time. We were researching one thesis for about 14 months. We met 140+ companies in that space to try to understand what is the right approach. Because we think if a company starts out two degrees separate at the beginning, if you launch a rocket, two degrees, you hit the moon or you miss the moon. So over the course of 10 years that's what we're really about, and we think that kind of investing setup lends itself to a market like this. The other thing that's really important to us is to be able to support companies across multiple rounds of financing, because the capital markets are so different than they once were. The next round is not necessarily guaranteed and we can have a level of depth in really understanding our companies and we can have the conviction to keep supporting them as they grow.
BP Great.
[music plays]
BP In honor of your early work experience– “What exactly is a container in J2E and how does it help?” Well, Johnny has the answer for you and has helped over 23,000 people over the last 11 years, so appreciate it, Johnny. As always, I am Ben Popper. I am the Director of Content here at Stack Overflow. You can find me on X @BenPopper. Hit us up with questions or suggestions for the show, podcast@stackoverflow.com. And if you liked it, leave us a rating and a review, because it really helps.
TT Thanks very much for inviting me on the show. My name is Tom Tunguz, General Partner and Founder at Theory Ventures. You can find me at tomtunguz.com.
BP All right, everybody. Thanks for listening, and we'll talk to you soon.
[outro music plays]