The Exchanges

Q When you invested in Andrel, what was your, you mentioned you had criteria and they checked it all. What was your mental, Oh, if I'm going to invest in a tech forward defense company, it needs to have X, Y, Z. What was that criteria?

A Yeah, so, um, Android all happened in a unique moment in time where Google had just shut down Maven, and defense had suddenly become very unpopular in Silicon Valley, and people were making arguments that ethically you shouldn't support the defense industry, and all this stuff that I thought was pretty ridiculous, because if you cared about Western values and you wanted to defend them, of course you needed defense tech. So, I started looking around to see who's building interesting things in defense, because if the big companies won't do it, then what a great opportunity for a startup, right? It seemed like a good, a good moment in time. And it felt like there was four or five things that you needed in order to build a next-gen defense tech company because there was a bunch of defense tech companies that just never worked or hit small scale. Number one is you needed a why now moment for the technology. What, what is shifting in technology that the incumbents can't just tack it on, right? Because the way the defense industry works is there's a handful of players called primes who sell directly to the DoD. And they subcontract out everything else, right? And if you're not a prime and you don't have a direct relationship, then you end up in a bad spot in terms of being able to really, well, win big programs and, you know, uh, survive as a company or succeed. Uh, so number one is w…

AI assessment note: “Number one is you needed a why now moment for the technology.”

Q Let's come back to outliers. So product market fit outliers. How do you identify an outlier team?

A Yeah, you know, uh, I think it really depends on the discipline or the area. For tech, I think it's very different than if you're looking in other areas. Um, for an early tech team, I almost used, like, this Apple framework of Jobs, Wozniak, and Cook, right? Steve Jobs and Steve Wozniak started Apple together. Um, Steve Jobs was known as somebody who really was great at setting the vision and direction, but also was just an amazing salesperson. And selling means selling employees to join you. It means raising money. It means selling your first customers. It's negotiating your supply chain. Those are all aspects of sales in some sense or negotiation. And so you need at least one person who can do that, unless you're just doing a consumer product that you threw out there, right? And it just grows, and then people join you because it's growing. Um, then you need somebody who can build stuff and build it in a uniquely good way, and that was Wozniak, right? The way that he was able to hack things together, drop chips from the original design of Apple devices, et cetera, was just considered legendary. And then as the thing starts working, you eventually need somebody like Tim Cook who can help scale the company, and so you could argue that was Sheryl Sandberg in the early days of Facebook who eventually came on as a hire and helped scale it, and Zuck was really the, um, sort of mixtu…

AI assessment note: “I almost used, like, this Apple framework of Jobs, Wozniak, and Cook”

Q You're around a lot of outlier CEOs, not only in the context of you know them, but you hang out with them. You spend a lot of time with them. What are sort of the common patterns that you've seen amongst them? Are there common patterns or is everybody completely unique? But I imagine that at the core, there's commonality.

A Yeah. You know, this is something I've been kind of riffing on lately, and I don't know if it's quite correct, but I think there's like two or three common patterns. I think pattern one is There are a set of people who are, and by the way, all these people are like incredibly smart, you know, incredibly insightful, uh, et cetera, right? Um, so they, they all, uh, have a few common things, um, but I do think there's two or three archetypes. I think one of them is just the people who are hyper-focused. They don't get involved with other businesses. They don't do a lot of angel investments. They don't, you know, do press junkets that don't make sense. They just stay on one track. And a version of that was, um, Travis from Uber. I knew him a little bit before Uber and I, you know, run into him once or twice since then, but like, um, he's, he was always incredibly focused. He used to be an amazing angel investor. I think he made great investments, but he stopped doing it with Uber and he just focused on Uber. Um, and as far as I know, he never sold secondary until he left the company, right? He was just hyper-focused on making it as successful as possible. So that's one class of ArchType. There's a second class, which I'd view as people who are, um, equally smart and driven, but a bit more polymathic may be the wrong word, but they just have very broad interests and they express tho…

AI assessment note: “I think there's like two or three common patterns. I think pattern one is”

Q Why do you think they're so relevant today?

A Why is YC still relevant today? Um, I think they've just done a great job of building sort of brand and longevity. Gary, who's, uh, taken over is fantastic, and so I think he brings a lot of that. Let's, let's go back to first principles and really implement YC the way that, you know, we think it can really succeed for the future. And I think they do a really good job, um, of two things. One is plugging people in, as mentioned, particularly your SaaS company, you want to have a bunch of customers instantly, your batch mates will help you with that. But also, um, it teaches people to ship fast, um, And to, to kind of force finding customers. And so because you're in this batch structure and you're meeting with your batch every week and you hear what everybody else is doing, you feel peer pressure to do it, but also it kind of shapes how you think about the world, what's important, what to work on. And so I think it's almost like a brainwashing program, right? Beyond everything else they do, which is great. It sets a timeline that you have to hit and it brainwashes you to think a certain way.

AI assessment note: “I think they've just done a great job of building sort of brand and longevity.”

Q Scaling a company often means scaling the CEO. What have you learned about the ways that successful CEOs scale themselves and things that get in the way?

A Yeah, I think it's two or three things. One is, um, figuring out who else you need to fill out your team with and how much can you trust them and all the rest. And so, one piece of it is very innovative founder CEOs always want to innovate and so they reinvent things that they shouldn't reinvent. Like sales is like effectively process engineering that's been worked through for Decades, you don't need to go reinvent sales. You know, you just hire a sales team and it'll work just fine. So one aspect is getting out of your own way on reinvention. There's certain things you want to rethink, but many of them you don't. Um, part of it is hiring people who are going to be effective in those roles and more effective than you might be. Often, you end up finding people who are complementary to you. Um, now that really breaks down during CEO succession, because what happens is often the CEO will promote the person who's their compliment as the next CEO, instead of finding somebody like them who can innovate and push on people and drive new products and new changes. And so often you see companies have a golden age under a founder and then decay. And the decay is because the founder promoted their lieutenant who was great at operations or whatever, but wasn't a great product thinker or technology vision like themselves. And so that's actually a failure mode for like longer term. Uh, related…

AI assessment note: “one aspect is getting out of your own way on reinvention.”

Q How do you think about that in an era of remote work?

A Remote work is generally not great for innovation unless you're truly in an online collaborative environment. And the funny thing is that when people talk about tech, they would always talk about how tech is the first thing that could go remote, because you can write code from anywhere, and you can contribute from anywhere. But that's true of every industry, right? You look at Hollywood. You could make a movie from anywhere, like you film it off-site anyhow, or on-site in different places. You could write a script from anywhere. You could edit the musical score from anywhere. You could edit the film from anywhere. You could write the script from anywhere. So why is everything clustered in Hollywood? Nobody would ever tell you, oh, don't go to Hollywood, go to Boise, and, you know, you could, you could work in the movie industry. Or finance. You could raise money from anywhere, come up with your trading strategy from anywhere, everything in finance is in a handful of locations. And so check is the same way, and it's because there's that aggregation of people, there's the people helping each other, sharing ideas, trading things informally, um, learning new distribution methods that kind of spread, learning new AI techniques that spread. There's money around it that funds it specifically so it's easier to raise money. There's people who've already done it before who can help you s…

AI assessment note: “Remote work is generally not great for innovation unless you're truly in an online”

Q When I think of a model, though, I, I don't think of an agent, I just think, well, I can't AI do that? Like, why do I need a specific What type of AI to book a flight to Mexico? Why can't ChatGPT just do it?

A ChatGPT in its current form, or at least in the simplest form, Um, is effectively interrogating a mix of like a logic engine and a knowledge corpus, right? It's like a thing that will look at what it knows and based on that provide you with some output. That's a little bit different from asking somebody to take an action. And that's similar to if I was, if I was talking to you and I said, hey, where's a nice place to go? And you didn't say, oh, you should go to Cabo or you should go to wherever, right? That's different for me saying, hey, could you get me there? Right? And you have to go to the computer and load up the website and book it for it. It's the same thing for AI, right? And so right now we have AIs that are very capable at understanding language, synthesizing it, manipulating it, but they don't have this remembrance of, um, all the steps that they've taken and will take. And so you need to overlay that as another system on top of it. And, um, you see this a lot in the way your brain works, right? You have different parts of your brain that are involved with vision and understanding it. Uh, you have different parts of your brain for language. You have different parts of your brain for empathy, right? You have mirror neurons that help you empathize with somebody or relate to them. So, uh, your brain is a bunch of modules strung together to be able to do all sorts of co…

AI assessment note: “That's a little bit different from asking somebody to take an action.”

Q Like where, you know, here's the, not the solution maybe to Alzheimer's, because that's like a, a big leap, but maybe it's like, you're not looking in the right area. You need to research in this area more. Like when is that sort of advancement coming?

A Yeah, I think it's a really good question. I mean, um, AI is already having some interesting advancements in biology, right? The Nobel Prize this past year in biology went to, um, Demis and a few other people who built, uh, predictive models using AI about how proteins will fold, right? And so I think it's already being recognized as something that's impacting the field at the point where it gets a Nobel. The hard part with certain aspects of biology and protein folding is a good counterexample. We actually have very good data. You had tens of thousands or maybe hundreds of thousands of crystal structures. You had solved structures for all these proteins, and you could use that to train the model.

AI assessment note: “AI is already having some interesting advancements in biology, right?”

Q this the other day, because I was driving, and you know, I got stopped by one of these people, and I was like, we're just teaching kids that, like, they don't even have to pay attention. They can look at their phone, the crossing guard's gonna save them, and then if the crossing guard's not, like, we're not developing ownership or agency in people. How do you think about that?

A I think it's, I think it's really, um, bad for society at scale. I mean, it's kind of like, um, there was a different wave of this, which was, you know, 1015 years ago with fragility and microaggressions, and everything can offend you, and you need to be super fragile, and all this stuff, right, which I think is very bad for, for kids, and I think that has a lot of mental health implications. Um, the wave we're in now, which is basically taking away independence, agency, risk-taking, I think that has some really bad downstream implications in terms of how people act, What they consider to be risky or not, and what that means about how they're going to act in life, and also their ability to actually function independently. So I, I agree. I think, I think all those things are things that we've accumulated over the last few decades that are probably quite negative.

AI assessment note: “I think it's, I think it's really, um, bad for society at scale.”

Q Do you think that'll be rapid, or do you, like, how will those breakthroughs happen?

A Yeah, it's kind of the same thing. You kind of, um, need to figure out what's the data set you're using. Uh, what kind of model and model architecture are using? Because different architectures seem to work better or worse for certain types of problems as well. Like the, the protein folding ones have three or four different types of models that often get mixed in, at least traditionally. A lot of them have moved to these transformer backbones, but then they're augmented by other things. Um, so it's a little bit of like, do you have, do you have enough and the right data? Do you have, um, the right model approach? And then can you just keep scaling it?

AI assessment note: “You kind of, um, need to figure out what's the data set”

Q Where do you see the bottlenecks today? And like what comes to mind for me are different aspects of AI. So you have, uh, from going all the way up the stack, you have electricity, you have compute, you have LLMs, you have data. Where do you see the bottlenecks being, where's the biggest bang for the buck? Like what's preventing this from going faster?

A You know, it's a really interesting question, and I think there's people who are better versed than I am in it, because there's this ongoing question of when does scaling run out for which of those things, right? When do we not have enough data to generate the next versions of models, or do we just use synthetic data, and will that be sufficient? Or how big of a training cluster can you actually get to economically? You know, how do you fine tune or post train a model? And at what point does that not yield as many results? Um, that said, each one of these things has its own scaling curves. Each one of these seems to still be working quite well. And then if you look at a lot of the new reasoning stuff that OpenAI and others have been working on, Google's been working on some stuff here as well. Um, when you talk to people who work on that, they feel that there's still, uh, enormous scaling loss for that still left, right? Because those are just brand new things that just rolled out. And so these sort of reasoning engines have their own big curve to climb as well. I think we're going to see two or three curves that are simultaneously continue to inflect.

AI assessment note: “Each one of these seems to still be working quite well.”

Q Like who, who's most likely to, to accrue some of the advantages of AI?

A Yeah, it's, it's, it's kind of hard because, um, AI is the only market where the more I learn, the less I know. And in every other market, the more I learn, the more I know, and the more predictive value, or the more I'm able to predict things, and I can't predict anything anymore, you know? I feel like every six months, things change over so rapidly. You know, fundamentally, there's a handful of companies in the market that are doing very well. Obviously, there's Google, there's Meta, there's OpenAI, there's Microsoft, Anthropic, and AWS, or Anthropic, um, x.ai, you know, Mastral has done some interesting things over time. So I think there's like a handful of companies that are the ones to watch, and the question is, how does this market evolve? Does it consolidate or not? Like, what happens?

AI assessment note: “there's a handful of companies in the market that are doing very well. Obviously, there's Google”

Q why most companies die from self-inflicted wounds, and what it really means to scale a company, and importantly, what it means to scale yourself. It's time to listen and learn. You've had a front row seat at some of the biggest, I would say surprises in a way, like Stripe, Coinbase, Airbnb, when they were just ideas. What was the moment where you recognized these were going to be outliers?

A So all three of those are very different examples to your point. I invested in Airbnb when it was probably around eight people. Stripe was probably around the same size. And then Coinbase only got involved with much later when it was a billion dollar plus company. And even then I thought there was enormous upside on it, which luckily has turned out to be the case. I think really the way I think about investing in general is that there's two dimensions that really matter. The first dimension is what people call product market fit, or is there a strong demand for whatever it is you're building? And then, um, secondarily I look at the team and I think most early stage people flip it. They look at the team first and how good is a founder. And obviously I've started two companies myself. I think the founder side is incredibly important and the talent side is incredibly important, but I've seen amazing people get crushed by terrible markets and I've seen reasonably mediocre teams do extremely well in what are very good markets. And so in general, I first asked, do I think there's a real need here? How's it differentiated? What's different about it? And then I dig into, like, are these people exceptional? How will they grow over time? You know, what are some of the characteristics of how they do things?

AI assessment note: “the way I think about investing in general is that there's two dimensions”

Q As an investor, what's the ROI on a 60 hundred billion dollar open source model? How do you think through what Facebook is trying to like do or accomplish? Is it just like, I don't want the competitors to get too far ahead. I need to absolutely.

A I don't know how Meta specifically is thinking about it. So I think I'd be sort of talking out of turn if I just made some stuff up. I think that in general, um, There's been all sorts of times where open source has been very important strategically for companies, and if you actually look at it, almost every single major open source company has had a giant, ah, institutional backer. IBM was the biggest funder of Linux in the nineties as a counterbalance to Microsoft. And the biggest funders of all the open source browsers are Apple and Google with WebKit. And you just go through technology wave after technology wave, and there's always a giant backer. And maybe the biggest counter to that is Bitcoin and all the crypto stuff. And you could argue that they're their own backer through the token, right? So Bitcoin financially effectively has fueled the development of Bitcoin. It's kind of paid for itself in some sense as an open source tool or open source sort of, um, form of money. You know, I don't know why AI would be different. I, a couple of years ago was trying to extrapolate who was the most likely party to be the funder of open source AI. And back then I thought it would be Amazon because at the time they didn't have a horse in the race like Microsoft and Google, or maybe it'd be Nvidia. And Meta was kind of on the list because of all the money they have, but in Prioris and…

AI assessment note: “IBM was the biggest funder of Linux in the nineties as a counterbalance to Microsoft.”

Q How do you go about constructing a better cluster? Like, if you take the presumption that the, the material that goes into my head, whether I'm reading, you know, that's one way, I'm conversing, I'm searching, how do I improve the information quality, uh, through a cluster or not, that my raw material is built on later?

A Yeah, I think it's a few things, and I think, um, Different people approach your processes in different ways. And this is back to the best people somehow tend to aggregate, or maybe best is the wrong word. There's a bunch of people with common characteristics, a subset of whom become very successful, that somehow repeatedly keep meeting each other quite young in the same geography. And again, it's happened throughout history. And so, A, there's clearly some attraction between these people to talking to each other and hanging out with each other and learning from each other, and sometimes you meet somebody and you're like, wow, I just learned a ton off of this person in, like, 30 minutes, and this was a great conversation, versus, okay, yeah, that was, that was nice to meet that person, they're nice, or whatever, you know, and, um, I feel like a lot of folks who end up doing really big interesting things just somehow meet or aggregate towards these other people and they all tell each other about each other and they hang out together and all the rest. And so I do think there's, um, sort of self attraction of these, these groups of people. Now the internet has helped create online versions of that. There's been a lot of talk now about these, um, IOI or gold medalist communities where people do like math or coding competitions or other things. Um, Scott, the CEO of Cognition, is a …

AI assessment note: “There's a bunch of people with common characteristics, a subset of whom become very successful”

Q Are there any historical parallels to anything that you can think of that map to artificial intelligence or AGI?

A Um, I think the thing that people misunderstand about artificial intelligence is that, you know, people are kind of viewing it as what you're selling as like a cool tool to help you with productivity or whatever it is. I think in a couple years we'll start thinking about it as we're selling units of cognition, right? We're selling bits of person time or person equivalent to do stuff for us. I'm gonna effectively hire 20 bot programmers to write code for me to build an app, or I'm gonna hire an AI accountant And I'm going to basically rent time off of this unit of cognition. On the digital side, it really is this shift from you're selling tools to you're selling effectively white collar work. On the robotic side, you'll probably have some form of like robot minutes or something. You'll probably end up with some either human form robots or other things that will be doing different forms of work on your behalf and, you know, potentially you buy these things or maybe you rent them. You know, it'll be interesting to see what, what business models emerge around it.

AI assessment note: “I think the thing that people misunderstand about artificial intelligence is that”