The Exchanges

Q many companies and so many startups, so many products, everyone's launching the speed at which companies and products are shipped is like thousand xing. How do you think about durability and moats when you look at a startup? Because a lot of founders get that question. Everyone's getting that question. How am I not a rapper? What are, what are signs that tell you this might be a durable thing?

A Well, I think there's two important ideas. One is something that, um, Jesse from Decagon said, which I love, and that is that moats are most often discovered, not designed. I think it's really easy. I've done this as a founder to get in your own head about like, Hey, I need a business plan that survives scrutiny from MBAs and VCs. I've got to have some really sophisticated, you know, idea of what my moat will be. And for that team, they just started shipping and it developed over time. Another great example of this is cursor. You know, they were criticized a lot for not having a moat, but it turned out that initially being a high NPS DAU product was really good. And over time they captured all the reasoning traces. They train their own models, the composer one, two models, and you know, so on and so forth. We know how that story plays out. So moats can be discovered. They don't have to be designed is one. And then I think the other is that we seem to have forgotten that the classic moats, none of the classic moats are based on how hard it is to make the software. You know, like we're not building self-driving cars. Most of us aren't. So it's network effects, it's scale advantages, it's brand effects, proprietary sort of data or what was historically called a cornered resource. Every moat from five years ago generally is still a good moat. We just need founders that have ambitio…

AI assessment note: “moats are most often discovered, not designed”

Q down their AI development. They paused Uh, their RL kind of phase on their latest model because of what they're seeing. So that's obviously a big concern for people, just how fast and smart these models get. Any thoughts on just that? That's like a big shift now instead of race ahead to the fastest, best model ever. Okay, we actually have to slow these things down. That feels crazy.

A I mean, without commenting on OpenAI specifically, I think that, um, maybe I'm a little skeptical on some of these things where I think the kind of the aura That Anthropic got from having a model that was too dangerous to release was extraordinary. And maybe they had a GPU shortage, you know, maybe it was, you know, the capabilities were more advanced than they actually wanted, you know, maybe they actually wanted to keep that proprietary model internal to extend their own lead. So I think there's a lot of sort of confounding factors that would cause you to pull back a little bit. Look, I do take the points about offensive cyber seriously. Which is we should harden all of our systems before we make them trivial to penetrate. Um, but I think the sort of concept of the model that's too dangerous to release it, it kind of conflates marketing, um, inference capacity, and then also economic considerations, like, do you want to externalize your competitive advantage or use it to make yourself better?

AI assessment note: “maybe I'm a little skeptical on some of these things”

Q Yeah, it's like you have to click three different matching shapes, which, like, come on, you can't do that? Hopefully by the time this comes out, I haven't seen it. I think, uh, it's a few people I guess in a row, I'm like, I haven't seen it yet. Oh, man. Okay, uh, favorite new AI product? I don't know, favorite AI product recently that you've, they've given me joy.

A Oh man, I've, I've spent a lot of time with the personal agents. I think Grok's okay. I'm going to give it to GrokBots only because it's just so unhinged for it to be so ambitious about sort of caching credentials and getting work done on your behalf. Like I love it and I expect it from a startup. Um, but they actually are doing things that I think no other sort of big company would do. It's really, really well done. It's a really thoughtful UI. It's got a really powerful foundation model. I think it also is like, okay, wait, maybe this is not a two horse race on the sort of model side. So I, I just think the product is like fun and ambitious and, uh, and is sort of taking risks that other products like that wouldn't take.

AI assessment note: “I'm going to give it to GrokBots only because it's just so unhinged”

Q tools aren't becoming the most productive person ever, you're going to become part of this permanent underclass and fall behind and have a really hard time. And some people are joke about it. I think a lot of people take this really seriously and stresses a lot of people out. How real of a concern do you think this really is? How seriously do you think people should take this?

A Not very seriously, and it's a funny dark fantasy that we seem to have as, you know, Silicon Valley collectively, like things have never been better, really by almost every measure, by how sort of distributed all the opportunities are, by the kind of technology we have access to, to the types of ambition we're allowed to have, and to the number of companies that are sort of independently working on things that are winning, and yet there's this sort of discussion of permanent underclass, being outside of the light cone, I've heard it, And it's not just something for deep insiders or outsiders. It feels like there's a real fear kind of from, you know, researchers at foundation model labs all the way through to the Silicon Valley layman. I mean, here's like a couple of points that I think are really important. So first, I think the last era of tech was a lot more centralized. If you look at network effects, that's sort of the gold standard. You worked on a network effects product. Um, that's the gold standard of businesses from the mobile era. And those things led to dramatic centralization, right? Of course, all of them are definitionally sort of the end of one networks. If you look at what's happening now, it's like every part of the stack, there's not even two relevant players. There's like, you know, you've got labs, you've got open weight, you've got different variations with…

AI assessment note: “Not very seriously, and it's a funny dark fantasy that we seem to have”

Q What's, what's one thing you've learned from each of those guys?

A They're such extraordinary leaders, founders. I mean, the thing that I actually feel so grateful to be a part of that they really embody to their core, I think is a feeling of stewardship for the technology industry and for really the country and the sort of maybe the Western way of living and thinking. Um, you know, and if you look at sort of a Ron Conway, you know, um, or, uh, Brooke Byers, Tom Perkins, There was this feeling, I think, of obligation, um, to sort of leave it better than you found it from an industry perspective, and that sort of aspiration goes way beyond just being the best investor in the world, though we want to do that too. I see them both show up that way over and over again, where they want to do hard, important things that don't directly benefit the firm, or at least not singularly, because they're just sort of important. You know, Ben has done a lot of that in the direction of the industry and Mark in sort of the direction of the country, though, of course, both work on both. Another really cool thing is just to see how I think Mark and Ben, but maybe the firm a little bit has shaped our collective ambition as a founder community. I think if you look at five or seven or 10 years ago, deep tech was deeply unpopular. You know, it wasn't a high status thing to be working on. It was very fringe. And, uh, and now it's become very popular, high status, and m…

AI assessment note: “Ben has done a lot of that in the direction of the industry and Mark”

Q the least valuable and engineering skills, the most important and valuable because they're so concrete. And it turns out that's what AI is the best at, the things that are verifiable and, and, you know, you know what success looks like. And so the question I think about now is just like, Which skills can you not just turn into a loop because the output is so hard to verify?

A Yeah, I think that's right. And I think that those are going to be the rate limiting factors because you can only do one steak dinner a night. I guess you could do a steak lunch, but you know, to some extent there's going to be these rate limiting factors in every system. Um, I think a useful way to think about it is anytime the model's making a mistake or doing something you wouldn't do, what do you know that it doesn't know? Um, and there's a really interesting example I heard from Ale at Kavak. He's probably the most sophisticated thinker on this stuff that I get to hang out with. Where he said anytime their agent, they have an agent per customer, they sell used cars online. And when the agent gets stuck, it actually calls a human and the human will coach the agent through. Now the magic of that is not only does it unblock the agent, but of course the agent then captures all the traces and learns from it. Like that's a really interesting mental model. It's either a knowledge gap or a data gap that you have to give the agent. And the next time it shouldn't have to call you.

AI assessment note: “a useful way to think about it is anytime the model's making a mistake”