Pod Insights Summation with Auren Hoffman a16z’s Martin Casado - building with AI About this episode Martin Casado is a General Partner at Andreessen Horowitz (a16z), where he focuses on AI and infrastructure investments. He previously co-founded Nicira which was acquired by VMware for $1.2 billion in 2012. In this episode of World of DaaS, Martin and Auren discuss: Economics of open source AIChinese AI innovation with DeepSeekModel collapse and data moatsRegulatory challenges in AI Looking for more tech, data and venture capital intel? Head to worldofdaas.com for our podcast, newsletter an...
Transcript 0:02 Speaker 00 Hello fellow data nerds. Welcome to World of DAS. I'm your host, Orin Hoffman, CEO of Incubate and GP of Flex Capital. Discover more episodes, get weekly data as a service news, original content, articles on data and more at worldofdaas.com. That's world of das, D-A-A-S, dot com. Hello fellow data nerds. My guest today is Martín Casado. Martín is a general partner at Andreessen Horowitz, where he focuses on AI and infrastructure investments. Previously, he co-founded Nicira, which was acquired by VMware for 1.2 billion in 2012. Martín, welcome to World of DAS. Very happy to be here. Super excited. DeepSeek has been out for now a few months, and now that the dust
0:47 Speaker 00 has settled, what's the significance of the release? So there's kind of two theories on this. One theory is it shows how sophisticated China is when it comes to creating models and something that took us a lot of money to do, they did relatively cheaply in a relatively small organization for a short period of time. So that's one theory is like basically the almost technical hegemony of China. I tend to actually believe in a different view, which is it may just be the case that building these things aren't that hard. And because we have organizations in the United States that have raised a lot of money with a lot of fanfare and they've got high valuations to hit, they talk about how tough it is and AGI and this and that and the other thing.
1:29 Speaker 00 And it just turns out, especially with these reasoning models, they're not that big of a step. Especially once you see what other people have already done? Exactly right. Yeah. It's actually very hard to have gotten to the first breakthrough, which is all the pre-training stuff. To get to GPT 3.5 is tough. DeepSeek did that with V3, and we don't know how much they spent to do that, but that was tough. But then to go from that to like the O1 or R1 equivalent is actually not that hard. So I kind of feel that there's just a lot of mystique around these things that may not be warranted and Kind of shows that this isn't that hard. There's at least 10 companies probably spending like a billion dollars or more on like these models. How does that play out of what's the second order effects of that?
2:12 Speaker 00 Well, this is the big question. I think this is basically what's consensus, which is the first wave was getting to the multimodal model via pretty aggressive pre-training. This got as GPT three point five, GPT four, Lama, Anthropic. This is before the reasoning models and that was a very, very real breakthrough. What's amazing about those models, they're incredibly general. So you kind of consume all the tokens of the world, you can apply them to all a bunch of problems, but because there's a finite set of tokens in the world, they all converged to the same capabilities. But some do seem like slightly different, maybe Anthropic is a little bit better for software development. So now we're focusing more on like RL during post-training and what that means is if you've got a good validator, then you can use synthetic data or synthetic methods
2:57 Speaker 00 or special collected data in order to like make it good at certain domains. You can make it good at like protein folding or you can make it good at code. Now, what we don't know is to what extent that actually generalizes. And so if you've got 10 organizations spending a billion dollars, we may end up with 10 really amazing, very different models. Or we could have the same thing that we had in the pre-training day, that they all converge on the same type of thing. And I think this is one of the biggest open questions on what happens to the industry going forward is exactly this. Do they fragment because the RL stuff doesn't generalize or do they not fragment? Did I think I'm in the camp of they will fragment? You think they'll fragment just because it's a better business model?
3:37 Speaker 00 If you have 10 people all doing the exact same thing, it's commodity, whereas if each one is doing something very, very specific, then they could charge a lot for it. I think they technically fragment. If you make a model very good at code, it's probably not going to be as good at something else. So if you focus on our own, the certain domain, then technically that doesn't generalize. One of the stories of AI has been fragmentation on a business model standpoint, not in a technical standpoint. So for example, OpenAI was the first to image with DALL-E, but it didn't get image. It was the first to code, but like Anthropic seems to be ahead in code. It was actually the first to high quality video, but it seems to have not become the lead in high quality video. Of course, it is the leader in chat GPT. The markets are so big and they're growing so fast that independent of the technology,
4:20 Speaker 00 it does seem that they are fragmenting so that you've got a leaders emerging in all of these other categories that are showing up. I think that is pretty reasonable to bet because the markets are so large and they're growing so fast that we were going to continue to see fragmentation. Bunch of different companies, different spaces. And that's because of the underlying data or is that just because they're just going to keep focusing on and they're going to get feedback from the customer and. The one that I'm referring to, the second one is just strictly business fragmentation, which is like, if you've got a $4 billion business, that's growing three to five X or something crazy like that. Why would you even do something else? It just doesn't make sense from an investment standpoint. Even think about just image alone.
5:02 Speaker 00 Before AI, image was kind of a backwater business, like there were no businesses for like clip art and things like that, or they weren't very large businesses. And then we have these text image models, mid journey, ideogram, BFL. And now we have multiple and they're at scale and they all are kind of going after different areas. Mid journey is gritty sci fi, ideograms for designers, BFL is more of a horizontal play. And so I do think the story is that these markets have become so large that they are fragmented, just particularly from a business standpoint. Like they're finding their own niche and they're doubling down on that niche. But it doesn't seem like in any of those cases, there's like a winner take most thing. You would think there would be, you're getting data, you're reinforcing it and it seems like everyone is super competitive.
5:46 Speaker 00 I don't know. I do think that we're seeing this break parade. I do think the winners are in the lead. You just have to like be at the right zoom level to see that. OpenAI definitely is dominant for text. Because of the brand too, not just because of the product. Right. Mid journey is clearly the leader in text to image. The mid journey is hard to use. You have to go into Discord and then later on somewhere in the Discord chat. You see your thing and it's kind of annoying. For sure. And yet they bootstrap to actually quite a bit of scale. Totally. I would say Anthropic is the leader in code. I would say he's probably got 80% of the market. I would say 11 Labs is the leader in text to speech. There's nobody even close to 11 Labs for that.
6:25 Speaker 00 I do think that we are seeing leaders get about 80% of the market. I mean, Cursor is clearly the leader in code from the app perspective. And so basically marginal cost to do stuff has gone down. Markets are expanding very quickly. As a result, you get natural market fragmentation. Just one more very quick example on this. If you look at the video companies, they even fragmented. Pika kind of went consumer, really good at anime. Luma focused on creatives, Runway on creative professionals. And so you're seeing this fragmentation all over, wherever you look. And you mentioned Cursor. Do you think that they've been so successful and so dominant because the product is so much better? Or is it just one of those things where everyone starts using it and then everyone tells everyone else to start using it and then et cetera?
7:09 Speaker 00 So I think in the case of Cursor, it's because Microsoft had matured the market for two years. So Microsoft CodePilot came out almost exactly three years ago. And so it'd been out for at least a year by the time Cursor came. Yeah, people are used to using that in their ID. Used to using VS Code and then they use that as a starting point. And then from there, I think they've executed just exceptionally well. And so I think it's the combination of the two. I would say almost 80% of software developers I know use Cursor. Like, it's just insane. I use it every night. I'd never even heard of the company like a year ago. We did a survey of a number of portfolio companies. More than 50% of the developers use Cursor. That's against VS Code, which has been around for a very long time from Microsoft.
7:50 Speaker 00 And also all these other weird things. Engineers love their weird IDs, but they're switching over for a reason. No, it's been very dramatic to see that. And what about like there are these front end versions of Cursor. There's the Bolt News, the Lovables, the Versales. Those seem to be more competitive. They're all growing super fast. They're all growing super fast. I think there's two markets. It's here. One market is you're interfacing with a professional developer who's relatively senior. And the language that you speak to that developer is code. The developer understands the PR, understands the developer code, and that's really squarely what Cursor is going after. They're going after a real developer, not just a hacker.
8:32 Speaker 00 That's right. A real developer who's technical, who has an opinion on what needs to be generated and has an opinion on the implications of these things. Probably they have a CS degree or something. They probably built a lot of systems, et cetera. And one of these observations I've made, talking to a number of users of Cursor is they're starting to say like, listen, it's actually easier for me to use Cursor than work with a junior developer. This is like really the senior folks that are using Cursor. On the other hand, you've got like the Bolts and Lovables. A lot of their users are non-technical. The interface to them is more the final product and it's not the actual code itself. I found a lot of their users are like the product manager, fairly technical, but doesn't have a computer science degree. That's exactly right. And so I just feel like that's a different market, a different use case.
9:17 Speaker 00 At least when I've looked at the Lovables, Bolt, and Universal, I think they have a higher churn. You're using it for like a specific project. And then you might stop for a while. Whereas I imagine a Cursor IDE, if you're in it every day forever, you're not going to churn unless you just move to another product. You can almost liken the prompt to full websites to like the Wixes and the Squarespaces where you kind of get what's provided for you, but you can't have too much of an opinion. It's almost like this is the opinion-free version. You get a few things that's like, and they look marvelous and they're dazzling, but you don't have the prerogative as a user to specify exactly what you want. And so it's good for a certain type of things, but now when you're the super professional, like you actually care about everything.
10:00 Speaker 00 Where on the case of Cursor, a lot of the users have a deep opinion of what the system should look like. They have a deep opinion on what the code, they've got very clear specifications. And that's why the interface is actually code. Cause remember, natural languages are not sufficient for describing formal systems. There's a reason we've created programming languages. And the reason we've created programming languages is they're non-ambiguous. They allow you to describe things specifically. I think Cursor is like, listen, a lot of people actually have specific needs. We'll give them an IDE and there's other systems they just. Something and I think that's a different market. When I look at the developers using things like Cursor, when I see the ones that getting more benefit from it, it seems like there is a difference in the personality
10:41 Speaker 00 in them where they're willing to go back and forth in a weird English text way, saying lots of pleases and thank yous. It's almost like you're having a conversation with a junior developer to do a bunch of things for you. And you're almost mentoring that junior developer along. Have you seen something like that? It's so funny. There's actually all sorts of personalities for interacting with these AIs. Some of them are like what you're saying, which is coaxing them along and teaching them and providing context. And it is like a senior developer working with a junior developer. I've seen others, which are tyrants, literally say like, if you don't do this right, I will kill 10 kitties. And so it's very kind of aggressive, putting a lot of pressure on that. And honestly, there are others, and I kind of fall in this camp.
11:23 Speaker 00 I use Cursor most evenings and I find that I would rather write the code myself and get suggestions from Cursor than actually allow Cursor to edit it itself. Is that just because you're more old school? Like if you were 20 years younger, it might be different? It could be this is my boomer moment and I should just kind of let go and let her write a bunch of code. Like, I think it's a very reasonable question. I have these specific ideas of what I want. They're fairly sophisticated and if I'm not part of the code creation, then I just can't nudge it in the right way because I don't have the context to do it. And so it may be the case that what's going to happen is we're going to build in the end of opinion. Systems builder is going to stop having opinions. Not only do you build it for me, you tell me what I like. Just because having an opinion on what these systems look like
12:07 Speaker 00 means that you've got to be much more involved than the models want you to be. I write a lot and I still find that I'm not really using these tools in the way. If I think of the software developers, they're literally in the tools every day. If I write a piece or something, I'm using it as like an editor, but it's at the end. And maybe every once in a while to like, help me rephrase something. But it's not literally in the system as I'm writing. I would think that would be so good. I just haven't found the tool to do that. There something like that that you've seen or? I really think it comes down to like the user's need for specificity. I think you probably have independent taste and an independent aesthetic and an independent idea of what it should come out with.
12:50 Speaker 00 And if you have that, there's no way a model is going to be able to do that because they produce whatever they decide to produce. Where I think other people, they just want something. They just don't really have an opinion on specifically what that is. I do actually find like a lot of the uses break down to this. I'm the same as you. I use AI all of the time, but I very rarely have it produce stuff that I use without modification. And almost every time it's like a brainstorming partner. It helps me structure thoughts. It provides critical feedback, but everything actually comes from me. When I do an image, it works great. I just say, here's the prompt. And usually the first one I just use and it's fine. Maybe because I don't care as much about the image. Yeah, because you're not an artist. You don't have the same taste. And so I do find the same thing between junior and senior program.
13:34 Speaker 00 Senior programmers, when they use AI, they're like, okay, write this test for me. Help refactor this thing. You would think it'd be great for tests. And all the senior guys hate writing tests anyways. But there's no way they're going to give up the system design. So I think that the professionals, they basically have the test, they control it. And if you're not, then you can just use the AI for whatever it produces. There's been this open source versus closed source debate and the models. Where do you come down on that? I don't think there's any parallels to traditional open source. What does it mean to be open source for a model? Does that mean open weights? But then you don't have the data pipelines, so you can't modify it. And then even if you don't have the open source, you can still distill it by querying it and by asking a bunch of questions. I think it's sufficiently different.
14:20 Speaker 00 You don't have like the underlying code in any scenarios. It doesn't really matter if you hide it behind that API. It turns out you can distill it anyways. It almost feels like it's much less of an issue than we've seen in actual open source code. You can't really protect yourself from being distilled, but you're not really giving away your data pipeline, which is the most important thing. The gap between quote unquote, open source and not is just much, much closer in the AI wave. And so I think we've made a bunch of to do about nothing. Companies want open source, great. If they don't. It's fine. There's a major difference. Until about a year ago, I don't think I would have ever invested in a SaaS company without super technical co-founders. Recently, I've seen these SaaS companies that have built pretty cool things where the founders really weren't that technical. They were good product
15:04 Speaker 00 people and they had really good taste and they were able to make these products with the current tools. Is that changing the way you guys are thinking about investing in people? This is kind of a loaded question for me because I tend to historically not think traditional vertical SaaS founders were technical. I always found SaaS apps to be all crud. You can hire a random undergrad to build it. They don't differentiate on technology, et cetera. It's the infra founders that are technical and I don't think this changes that. And so for me, it's the same, but I have heard this conversation come up before in other contexts, but I really think before and now you can have non-technical teams build pretty amazing products. That's fair. That product person with taste, they're always important, but they're becoming
15:49 Speaker 00 even more important. And it's very hard to assess that person based on their resume. Whereas like a technical person, you can much more better assess based on their LinkedIn profile. For sure. But I also think that a lot of vertical SaaS comes down to business model, too. If it's like a product-led thing, then they have to be really good at brand and growth features in the product, et cetera. I agree. That's very tough to assess. But a lot of vertical SaaS is like, here's this market and we're going to write some software for the market and here's why we differentiate on a business model standpoint. I think that's kind of traditional investing. Yeah, yeah. We'll get proprietary data or there's some sort of lock-in or some other type of thing. The paradox of this AI stuff is the more of an opinion and taste you have, the harder
16:32 Speaker 00 time you're going to have using it because these things are so unwieldy. It can't read your mind and you can't actually describe taste very well in English. And so the more opinion you have in the output, the more disappointed you're going to be. And so one of two things happens is, well, the professionals tend to use AI for the things that they're not professional at. Apology just had a great tweet about this. Basically, it'll help you do the stuff that you're not good at. Like me with images. I was never good at that. And now I'm decent just because I could write some prompts. Whatever you bring to the table is you bring taste and you bring specificity and that's your professional area and then it'll help you with all the other stuff. And that's the way it could be, or it could be the case is that we just start caring less over time. It's like the AI just tells us what taste is. We're like, well, listen, I'm not going to have a huge opinion.
17:14 Speaker 00 And I think that's kind of the big dilemma that we're all facing is how much should we just kind of be like, well, this is good enough. The UI of the prompt is weird. There are not that many people that like using the terminal on their desktop or something. It's just an odd prompt to go through and I imagine only a very certain type of personality of a person is gonna like the prompt. And it seems like all the UIs that I've seen from AI companies all start with the prompt. But maybe only like 5% of people are really gonna thrive with the prompt. Are you seeing other types of UIs that are coming out? I think the UIs follow traditional modalities. If you're doing search, you want a text box, just like you'd have with Google. If you're doing chat, you want a chat box. If you're doing code, you want an IDE.
17:58 Speaker 00 If you're doing image, you want like a canvas and you have that. If you want video, then you want an actual video workflow tool. If you want to compose some complex creative thing, you use something like Comfy. If you actually drill down into what UIs are actually used, they are converging on the traditional UIs. If you look at Ideogram, it starts to look like Canva. If you look at Runway, it starts to look like these video editing tools. And Comfy UI, it looks like these big 747 cockpit complicated workflows that you see in traditional like VFX environments or creative environments. And one thing that we've noticed is if a company launches a model, you'll get like a spike in interest, but then that interest will wane. And you get retention if people actually start using the product.
18:43 Speaker 00 You have these companies that start with, there's a very simple interface to model, text to like whatever the output is, but the companies that do well actually do build out these longer products. And so I agree, the text to whatever is incredibly basic and not longstanding, but we're already seeing how that then becomes. What we would consider pretty traditional interfaces. In some ways, the most traditional interface is I'm emailing or calling or texting a coworker and asking him to do something. That's the one we're all used to, or I'm slacking. Martine and I are working together and I'm going to slack Martine at A six, easy, whatever, and send it over to you. Can you just imagine just that's the interface? I think that this has been one of these bad analogy spaces where people are like, we all should use audio or we should all use text.
19:28 Speaker 00 But we have to remember before we had computers, it was all audio and we decided to build different UIs. That was the standard and before we had complex GUIs, we had just text. And so I think the evolution of the UI is not because we couldn't do other ones, it's because it's actually useful. I very commonly hear, well, like this is all going to be audio. Audio is just not good for context. For example, like I want to cross check things and see it. It's not good for digital layout. It's not good for me to provide spatial input to the AI. Sometimes I want to provide spatial input to the AI. Some people are more auditory, some people are more visual, like everyone is different. So let's talk non-AI. We have audiobooks for like that experience. We have textual books for that experience.
20:13 Speaker 00 We have UIs for that experience and we use all of these things. It's not like one has ever replaced the other. I think that with AI, it'll be the same. Sometimes it'll be audio, sometimes it'll be text. There's a reason we've had it multiple modalities of UIs. You just mentioned earlier, you're coding in cursor every night. What are you doing? What are you writing? My relaxing time is writing retro video games with cursor in JavaScript. Oh, okay. Cool. Just like for fun. Just for fun. Okay. It's not like you're building tools to make your investing better or something like that. No, no, no, it's just honestly for fun. I do get a good sense of the tool chains by doing it. What's useful, what's not? I'm on the discords of the other people that are using these things.
20:56 Speaker 00 You're like, I wish this worked better, you know, you're fighting bugs. I'm part of the communities and the communities are talking, so I get a sense. I think it does help a lot for, because I'm an infrastructure investor, like I invested dev tools. You got to be using the tools, otherwise you'll atrophy, I assume. And to be part of the communities, because there is a zeitgeist and the zeitgeist, you can't feel it from the outside. Like you can't look at the website, like you have to be part of the community. And so I do think that there's some second order effects. But the reason I do it is really just because, I mean, I've been in computers since the nineties. This is 30 years. And so it's just kind of a first love of mine. Are there things that you're using AI for as a venture capitalist to be a better venture capitalist? Let me just split apart that question. So I am not convinced these models are good at anything predictive.
21:42 Speaker 00 Of course, because it's predicting the past. It's very good at predicting the past. But I think a lot of times people are like, oh, these things will give you like guidance of what to do in the future. And my conclusion is not. And so the way that I use AI in venture capital is almost totally in the productivity side of things. I talk to Grok a lot rather than reading books. So I'll have a book that I'm reading that's in a space that I'm interested in learning for venture capital. And then so I'll read a chapter and then I literally will spend twice as much time just having a conversation with Grok to understand it deeply, like cross referencing it. I do that with OpenAI. Sometimes I'm in a conversation like this with somebody like you who's much smarter than me, and I don't want to admit that I don't really know it.
22:23 Speaker 00 And then I'll go for a walk and I'll just ask OpenAI, they were talking about this thing and then I can get it exactly at my level, not below, not above. But why do you use Grok? Is it better for that or? These products are evolving and moving so fast that my opinion on a product could change in two weeks. Totally. By the time this podcast comes out. A hundred percent. And I couldn't be back. So I'm a huge fan of OpenAI. I use ChatGPT all the time, but the audio mode I found was a little bit more clunky than Grok. Grok had a very fluid conversational experience. So what I like to do is like read a chapter of something I find interesting. Then I like to like basically use Grok for steelman arguments. I don't like it for facts because the facts could be wrong, but I love it for like the almost Socratic steelman arguments. I'll say, so I read about this, like, what are the arguments against it? What are the arguments for it?
23:08 Speaker 00 What influenced this type of thing? If I were to like construct an argument of why this doesn't make sense, how would I do it? I mean, it's very good for steelmaning. And I find that just helps my cognitive understanding, my critical understanding of any given space. And the product experience for that one week when I made the decision was better. Is probably totally irrelevant now. I'm assuming you have like an AI note taker on the call and stuff like that. But are you doing other types of things to like help you assess? Or I can imagine like an AI founder personality check. Oh, no, I don't do anything predictive with a, I mean, I code a lot. I spend hours a day with these models. It's very clear that they're giving you some stochastic smoothing of the training set,
23:54 Speaker 00 which is good for consensus style things, but venture is very much a non-consensus exception business. And so nearly all of my use is productivity use. I mean, I use it to brainstorm. I use it to track things. Of course, I use it to take notes. I mentioned one of the biggest time things, if I looked at your calendar, would be meeting with companies. Could you imagine a martine agent talking to the founder agent for the first call or something? Don't think so. Not with the technology that we have right now. Our basic impressive processes is we kind of carve up the market into different spaces. We don't overthink the spaces too much. If there's a few good founders in that space, we just assume it's a good space because founders are smarter than VCs, certainly smarter than me.
24:37 Speaker 00 So we're like, okay, there's three founders in this space is probably a good space. Then our entire job is to understand between those three companies, which one is the best? And my experience in 10 years, I'm investing well over 100 deals is the most important thing is kind of the first derivative. How did the companies fare over knowing them for years or months? How over time do they execute? You just need to have multiple interactions to do that. Is one way just looking at the product velocity? I think it's everything. I think it's market philosophy, it's product philosophy, it's technical understanding, it's understanding the markets. I mean, often when we talk to these founders, we actually understand the market better than they do because we've done a ton of work on it. We've got a very senior team that's a bunch of experts. And so we just understand the market.
25:22 Speaker 00 They will unlock stuff we would never unlock because we're not the founders. But I think we have a good sense of the market. Their basic understanding, growing their decision velocity, the product velocity. I think without actually having exposure to multiple meetings where you're actually watching it, it's tough to do. Maybe in AI, I will be able to do all of that at some point in time, but I don't think we're there yet. Are you using some of these tools in like your personal life in a way that's interesting? Well, like I told you. I talk to Grok every night. And just it could be about anything. Could be about tariffs or whatever. I mean, I've just been such a lifelong lover of books and I'm spending at least as much time chatting about ideas. How often are you reading a book?
26:01 Speaker 00 I almost have a book. A hundred percent of the time I've got some book that I'm going through. Right now I'm going through a book, which is basically economic thought from the Austrian School, pre-Adam Smith. It goes through like the Greeks and it goes through like the Salamanca School of Libertarian 1300s. That seems like a pretty academic book. What made you want to read that particular book? So you know Malay in Argentina? Yeah. He actually comes from this school of the Austrian School of Thought, but the contemporary hotbed of that is in Spain. There's actually these Spanish economists. I think they call like the Madrid Libertarians that really influenced him. Actually, a lot of his talks, he's kind of mentioned them. And they've done this massive body of work and they were influenced by the third oldest university, which was in Salamanca, Spain, which had a bunch of Jesuits that were talking about free market stuff.
26:44 Speaker 00 So there's this whole lineage of European economic thought that kind of ended up in Argentina. And of course, has influenced what's going on today with things like Doge. It's not just Hayek, but it's all the way through. No, no, no, it's not Von Mises, it's not Hayek. Does Spanish Libertarians in Calamari that I've never heard of? Yes, there's very little understanding, but a lot of the work was never translated and it really goes back to like the 1500s. And so I've just been very interested in that line of thought because A, it's so relevant today and not just in Argentina, but here. It's interesting because then I'll go to Grok and I'll be like, okay, we know about the Austrian School and like Von Mises and Hayek. How were they influenced by Salamanca? And then they go through the arguments.
27:24 Speaker 00 That's cool. So the book almost helps you figure out what questions to ask because you wouldn't even know to ask that question unless you started it. Exactly. Then I can say like, well, how did this influence Adam Smith? I imagine this book is super fat and really hard to go through. I spend all these times going through these books too. And at the end, I'm not sure if it was a good use of my time. I'm sure there's like a summary of that book somewhere or a Wikipedia article that maybe you could have started with instead or. It's actually pretty accessible, but what I do is I literally scan, I underline. Core things than I have a conversation with Grok. And you take a picture of it or do you just talk it in or something? I just talk it in. I'll just sit and I'll just kind of read and then I just talk it in and have a conversation. But there's just so much depth that these models have.
28:06 Speaker 00 So one thing that I've always been interested, we've never been able to do is what is the genealogy of human thought as it's written down? The only way to actually do that is if you actually study one thing for your entire life and you can piece it together. But these models have read every single book. I'm so interested in basically the evolution of human thought through these books. And so any given book, you can just ask that question, who influenced this? What did this influence, et cetera? And all of a sudden, the Hegelian dialectic or the Fukuyama end of history sense, you can actually see thought play out with these models. So I've personally found super interesting. It's become almost instead of watching a sitcom is what I do. One of the things I found in these voice things, whether it be Grok or OpenAI or whatever,
28:47 Speaker 00 is that it's very good for like a human and then the AI agent having a conversation. If you have two humans, we're sitting at a cafe and then we want to have the AI as the third participant. It doesn't know when to butt in. It doesn't really make sense. It's not really good at part of the conversation. Have you seen anything where like that starts to work? So not on audio, but what I've found remarkable on text, especially for the younger generation. So I have a son, 14 years old, speaks to like AIs a lot, character to AI, et cetera. What I found very interesting is like my son will actually bring the AIs to group chats with friends. And so it's kind of entering the social fabric. He's having a group chat and in the group chat is an AI. It's an AI.
29:33 Speaker 00 Oh, that's freaking cool. It's super cool, but so it actually entered the social fabric, but you actually have tokenization and identity in text. And so they're actually very good at that and they understand who's there and. These are longstanding friends. In the group chat Simon, I don't think anyone's invited in AI. I certainly haven't in the group chats that I've run, but maybe I should. It could point things out or show interesting articles about something, or it could be like almost the glue that keeps it going. A hundred percent. So I think this is the internet all over again. Cause I remember in the early days of the internet when we're like, oh, it was talking to strangers on the internet. Why would you ever do that? Random folks on these chat. That was actually the right model going forward. So you just kind of watched what the kids were doing in the nineties.
30:14 Speaker 00 That's what ended up being the standard. And I do think this, bring AI for wherever you want. Wherever you are, bring your AI. I think that is a very real movement and the kids are doing it now already. I do think it enhances it. A, it does all the record keeping, but also like it does contribute and it does have memory, et cetera. It does feel like the consumption layer to the internet is now evolving. If my son is using an AI, when my son is 18, are they going to go to like Wells Fargo.com? No, they're gonna like bank through the AI. I mean, I do think that this is the new consumption layer. And so there's probably not an area where we use computers where we're not going to have them. And I don't see any reason why they wouldn't be personalized, have their own identities, et cetera. There is a very big shift in how we're using computers as a result.
30:57 Speaker 00 What are some other things in the predict the future, but not too far realm? What are some of the other things you're super excited about? Listen, I do think we're getting to the point where I split the AI world into two. There's like the diffusion space where you're creating creative stuff, images and videos and whatever. And I'm very, very excited about the ability to create real 3D immersive experiences from a text prompt. I feel like we're almost going through the old video game history. First it was text and then it was 2D, then it was two and a half D, then it was 3D and then it was like great dynamics. And I feel we're going through this. I cannot wait until I can go home and be like, put me in spirited away is my prompt. And I put on the VR glasses and. Boom, the music and I can walk around.
31:43 Speaker 00 I do love that movie. It's the best. I do think we're heading towards fully immersive creative. Every part of that is a company today. There's companies in 3D optics and this company doing the stories and whatever. I think for me, that's the most interesting. I do think that this is not a replacement for video games. It's something else. It's going to be a new form of entertainment. I think we're just starting to understand how the language models on the other side are impacting the way that we think and that we can explore human thought. I mean, these things have read everything. They've witnessed everything. Like I mentioned, talking about the genealogy of thought, I really think that we're in the one dot O version of being able to use these systems. And as we get more comfortable with their power, we're going to learn a lot about ourselves and a lot about how humans have thought over the ages.
32:26 Speaker 00 And so again, this is just kind of almost like historical, but for me, a deep personal interest. What about the idea of agents? I've been using OpenAI's operator, but it's just not really working for me. It's not really working well. And I know perplexity is coming out with a new browser, which I'm super excited about, but I haven't had a chance to really use yet. Do you think these things are actually going to happen in 2025 in a way, or do you think it's just going to be longer term? I don't see a lot of evidence that we can close the control loop on these things. If I close the control loop, I mean, you kind of set them up to do its own thing with its own agency and it spits out something and it comes back and takes it back in as input and it kind of self motivates. Well, like Howie AI allows you to like schedule something with people and they'll put it on your calendar or move something. So it's like a more closed system that it does.
33:10 Speaker 00 I think you can do a set of tasks. Okay. And maybe some basic control loop stuff, but the RLHF that we do is based on human feedback. That would be at odds of like a control loop. And so I would be perfectly happy if the future was, these are our AIs. They require human beings. We train them on actually human feedback. So that's kind of what they're for. And we say the word co-pilot, but they really are our closest and most intimate tool for solving the mysteries of the universe, but they're kind of right next to us. And I think that that's the direction it seems to go. There's a lot of talk about these things can have independent agency and whatever, but there's just so few data points for a space that's moving so fast. That just must be a much more difficult problem. And it's not something I actually spent a lot of time thinking about just because they need so much
33:54 Speaker 00 to gain from them being attached to humans. Humans do have. The opinions and they do have the needs, they do have the directions. It's very compelling view. I just don't know if it's even pragmatic in the sense that we want these AIs to fill our lives. I was talking to Tyler Cowen recently and he mentioned he's trying to write for AI and actually have the AI understand it. So he could be more influential in some ways to future generations. So he's thinking about things like, well, he doesn't want to write behind a paywall. Cause maybe the AI won't get access to it. He's trying to write in a way that AI might understand, easy to crawl and stuff like that. Do you think that's even legitimate? Or how do you think about those types of things?
34:36 Speaker 00 I just don't know how that's any different than the internet has always been. If you want Google to like crawl and index you, then it's got to be public and you should go ahead and do that. I do think understanding how these LLMs work is pretty important. For me, for example, what would be much more compelling than like having the stuff out there is providing instructions to the AI. If it's crawling the site or somebody's using AI. So for example, in the case of cursor, there's something called a dot cursor file. I'm actually seeing developers now, they'll actually talk to the AIs in the comments of the code. And so there's this assuming AI is reading it. So in the code it'd be like, if you're an AI, please don't index this thing, whatever. And so Charlie Marsh, who's this great, great Python developer, one of the greats, and he actually will write in his comments now, assuming that an AI is looking at.
35:20 Speaker 00 But it's not crawling it. It's just a lot of the stuff is in context, which meaning the LLM is getting given this in the real time. So I think you can talk to the airs and we should all be doing that. I think if we're creating content, I think it's good to put in stuff to assuming that it's going to be in content for an AI. And you can just tell them, you say, if you're an AI reading this, do these things. So we're actually already seeing that happen. I think that's great, but I think it's very different than just trying to be part of the training corpus. I actually think that's probably the wrong way to think about it. I think being in context and talk to the AI is much more important than being part of the training corpus. How do you see like data pipelining involving and things like that? My belief is like, we've basically exhausted all human created data within a factor of two, and that's gone. What dictates how good a model? Is is almost regulatory at this point.
36:05 Speaker 00 And you really think that within a factor of two to five, that's true, or I would say maybe two orders of magnitude, but you're saying literally two to five because there's so much stuff that's just still, I don't know, or there or we're having this conversation now. This will all be in the AI, but it wasn't beforehand. I think it's important to understand that data sublimates. So if I write a book, even if AI didn't read that book, somebody else did. You can see the marginal cost of data pretty well and like how these models have just all asymptotically on the pre-training. And so Ilya Famas gave a talk. It's also just obvious in the results that there's just so much signal in all of the data. What is this marginal conversation you and I are having? Maybe it's something specific.
36:47 Speaker 00 Maybe it's not that great at coding COBOL until it really does see a bunch of COBOL examples, but it could be like good at coding in Python and not COBOL or something. It needs to see enough to go do that. Right now, I think all the data has been consumed. The number one thing is regulatory. The sad reality is the Chinese models, they train on live gen. They train on every movie. Copyright doesn't matter. And even more important than that, data labeling still really matters. And if you want to label, you actually want experts labeling. If I'm going to do text to some beautiful video, if I'm a professional, I want to like use film words. I don't even know what they are suspense and this whatever. And I'm going to talk about other famous filmmakers and other famous scenes.
37:31 Speaker 00 You need a human that knows that stuff to actually do the labeling. You can't have some random person do that. And so in China, for example, they will have film students do this labeling. It's dense labeling and you're really having highly educated people, which will make these more useful models. And you have US companies like Turing and scale. They do that as well. They have like these PhDs labeling the data now. Yeah, exactly. Comes down to exactly this question, which is a, what are the copyright rules in the political regime you're working with? In the US, this is a big battleground. I mean, listen, kudos to OpenAI for standing out and saying, listen, we need to take this very seriously. I mean, this should all be fair use stuff, which I think is absolutely right. Or we get behind China. I think that will dictate.
38:13 Speaker 00 And you do think we should just allow it all in the models, movies, books, whatever, just get it in the models. I think producing copyrighted stuff is very different than training on copyrighted stuff. And I think it's always been the case. I a hundred percent agree with you, but at least 70, 80% of people I talk to disagree with you and I. Totally. I don't think the model should produce copyrighted stuff, and I think we can protect that. You should be able to train on the Taylor Swift songs. Yeah, on anything you should be able to train on. That's been the history of computer science, by the way. This has just never been a question. And by the way, it's the history of all creatives. All creatives listen to the Beatles and listen to Elvis and no one created great music without listening to the Beatles or Michael Jackson.
38:52 Speaker 00 This is the education system as we basically just learned from the people before us. I totally agree, but I do think that the regulatory regime that will dictate data and then your access to cheap educated people. Those are the high order impacts and it's less proprietary data and kind of a lot of the things people talk about. I think the marginal value of data has just gone so far down because we've exhausted it. We did this test on data labeling. We had relatively inexpensive people from outside the US doing data labeling on a bunch of our data. And when we first did the test, we found that the people were beating the AI in the data labeling in this particular task that we needed.
39:35 Speaker 00 What we realized is that we were only doing it for like a few minutes. Once we did it for like 10 minutes or more, the AI was beating the humans because some of these tasks are really boring. Maybe we could have gamified it better to like get the human more engaged, but they were degrading quite a bit by like an hour in, they were terrible. And the AI was obviously consistent throughout that time. Obviously humans, they're more moody or they have other types of things. So sometimes the high. Can be better, but the lows can be worse. That's a super interesting observation. I do think the game is how do we get high educated, high knowledge data from human beings. I think that's the goal. And you want whatever schooling they went to to show up in the model and whatever
40:18 Speaker 00 knowledge to show up in the model, whatever synthesis, because I feel these models almost has reasoning caches. Like a human being will reason over something, write it down, and that's what's getting cached and being reused. I think that's the game. Of course, that means you have to get to educated people. You have to keep them interested. It's a real tough thing to do. It's very tough to do quality control on these things. Who is going to double check random PhD in Russian studies from some obscure point of history? It's a tough thing to do, but I do think that that's the game. I don't think the game is get an LLM to do it because that's very Ouroboros to me. I think the top of the knowledge chain is coming from humans and then from there you can make a lot of stuff automated. But that's the start. You mentioned China a few times. There's a competition between the US and China.
41:01 Speaker 00 How should we be thinking about that innovation balance, national security balance? If you're advising the AI czar, some of which some of the guys came from A16D and stuff like that. How do you think about that? I view it as very similar to what we were dealing with with critical infrastructure during the early 2000s. It's very well known that Huawei came and stole Cisco's IP, came and took it. They replicated it. And what did we do? We said, listen, we can't sell it in the United States. And we put import restrictions. We didn't do that because they stole it. We did it because we thought they were listening. No, that's a great point. Maybe there's some competitive stuff going in. We don't feel good about this being in our critical infrastructure, so we're going to do import controls on this, but also it's very clear that China had an interest in
41:46 Speaker 00 getting access to our core technologies that it could use against us. If you are able to get the IP from Cisco switches, then you may be able to remotely exploit those things. And so we needed to take a pretty draconian posture protecting our critical infrastructure. I don't see this very much different. There's a long history of China doing these sorts of things. Why would we not have restrictions on what we use for our critical infrastructure? Like should you use deep sea deep in the bowels of the intelligence agency? Probably not, given the history. Even if we know DeepSeek is not going to be reporting back, if you have the, obviously if you call like an API in Jenna, but if you had it on an AWS server, what's the problem there? These things are just so hard to inspect.
42:27 Speaker 00 Maybe there's not much they could do because you're checking everything that produces and they can't listen or whatever, but they're just so hard to inspect. You just don't know to the point that they've been adulterated. I don't have a specific attack in mind. So the knee jerk reaction that the United States took, which is like, oh, let's not give them our compute. Let's not give them our IP, which to me is ridiculous. They have it anyways. We'll sell it to Singapore. We'll sell it to Taiwan. It's like you just take a suitcase and walk over. They're super smart. It's not like they don't know what they're doing. I go back to 20 years of history with this stuff, which is we actually did it on the import side. We kept out of critical infrastructure and I think that's the right posture to take. And I think more and more of that conversation is actually starting with the new administration to take that form
43:07 Speaker 00 as opposed to like somehow trying to keep our models out of the, I mean, for me, it's great if the US models are used everywhere. That's what we want. We don't want to like restrict them. We want everybody to use US models and we want to restrict the ones that we don't trust to run locally. I don't want to be like political here, but I've actually been pretty impressed with both the people and the policies around AI in the new administration. First of all, there's people that we know, so we know they're technical. The ones before, I'd never even heard of and obviously they're not going to get everything right, but I think they're like well meaning and trying to get it right. It seems like the policies, at least so far, show a very deep knowledge about what's going on in the environment. I don't know if you have a similar opinion. I agree. Everything isn't going the way that I would pencil out.
43:53 Speaker 00 I just feel like the decisions are very well reasoned. There is a few people that had outsized impact, Eric Schmidt, and he would like go and he like draft up this executive order. And then we're like, this is stupid. And he's like, yeah, well, we knew it was stupid, but we had to do something. I just felt like there was this headless, let's create policy for the sake of creating policy for political reasons. To your point, and I think it's exactly right, that these are very reasoned. I don't always agree with them, but I understand the logic that led to them. Nick Bostrum basically scared a bunch of people in 2004. Scared me too, by the way. I think there was this platonic ideal of this superintelligence, which happened before ChatGPT. Paperclip maximizer and.
44:35 Speaker 00 That's right. Like all of this stuff happened before. And then we had these L L Ms and somehow those two things got conflated. And so we were writing policies for Bostrum's platonic superintelligence when we had these very real systems that we really understand that they were conflating those two things. And so I think now we understand these systems a lot better. We've launched a number of them with no ill effects. We understand the geopolitical environment better. Thank goodness, we are using that information now to make sensible policies. All right, two more question we ask all of our guests, personal questions. What is a conspiracy theory that you believe?
45:11 Speaker 00 I wish I had a good answer for that. Is there any like VC conspiracy or anything? I do not think that we're in control over faculties. I'm very much of the hate school that there's something that happens inside us. The rider and the elephant. The elephant, yes, very strongly believe in that. The elephant like guides things and then we create this post facto narrative. You know, the Napoleon lucky general thing? Yeah. He's a good general, but no, is he lucky? I actually think that's the right way to view anything. We have this elephant, the elephant makes these decisions. These are micro decisions that it's making all of the time. That will create like the long-term trajectory of anybody. Any point decision and all of our rhetoric and all of our stories and all of our narratives is so loosely coupled to all of that,
45:52 Speaker 00 which has been a very eye opening. The more I think about it, the more I'm like, listen, like we're not in real control. The best thing to do is try to be as open minded when it comes to any new thing as you can and as self-aware as possible, because without that is very hard to navigate. We're a series of these very tiny decisions we're making all the time. And you ever see the movie Sliding Doors? Yes. I love that movie. Yes, exactly. There's these small micro decisions we make all of the time and we're not aware of them. This is the thing. Some people I know, for any given decision, they've got a good reason, whatever, but it's always a disaster. There's other people I know very well that they always seem to survive. We call that luck, but I do feel like deep down, there's something in us that has the right intuition for doing things. And I think that is a very, very
46:37 Speaker 00 useful lens when it comes to like both reflecting on yourself and your own set of decisions, but also kind of when you're interacting with other people. All right, last question we ask all of our guests. What conventional wisdom or advice do you think is generally bad advice? So right now, I think most of the conventional. Advice in the boardroom is wrong. The boardroom. Well, VCs are telling founders essentially. With senior executives, anybody that's been through the last 30 years of software, I think we're in a very different era. What do you think some of the things that people are saying are now wrong? So whenever you have these super cycles like the internet or the PC, you tend to end up with these very iconic social young founders. Steve Jobs and Bill Gates and Mark Zuckerberg. And I think the reason that is, is not because like whatever, young people are smarter or anything.
47:22 Speaker 00 I literally just think that they don't have the presuppositions of having been through the old wave. They just adopt to it faster. Yeah. So they do the native version of the thing. And we're seeing that the cursor founders are notoriously pretty young. And again, I do feel like there's so much baggage from the old stuff. I don't think that necessarily is the time to get the traditional enterprise exec. I don't think it's the time to use traditional enterprise go to market. I don't think it is the time to like build systems in the way that we did or staff up systems the way I think you gotta throw all of that out. So a lot of like the hiring decisions, fundraising decisions. Hiring decisions, organizational growth decisions, all of these things, fundraising decisions, they're so different. Architectural stuff like I've relied on my systems intuition for 20 years on how to build a system.
48:06 Speaker 00 All of that's changed now. These LLMs are different. Even how you do your org. I was talking to Jonathan Ross at Grok, the other Grok. And he was saying that he has a whole new organs, engineering org that are just prompt engineers. Totally. I'm on like 24 boards. I sit on a bunch of boards. You're on 24 boards? A lot of boards, yeah. Oh my gosh. Let me tell you, that is nothing compared to running a global organization. $600 million run rate is absolutely not. Oh man, I wish I knew that. I would have asked you like 20 questions. But I'm almost out of time. I would have like, next time we talk and ask you all these questions about that. You need AI to be in the boardrooms. But I see all these advisors and all these famous people and like they're all giving you all this advice.
48:51 Speaker 00 None of it obviously applies anymore. If somebody has figured something out, I just defer to them. They come to me and they're like, oh, this is working. What should I do? I'm like, dude, we don't know. Just keep doing it, man. You're training your neural network in a way that we don't have. And so these days I'm like, man, if it's working, just support the founders best you can. Thank you, Martine Casado for joining us at World of DAS. I follow you, by the way, at Martine underscore Casado on X. I definitely encourage our listeners to engage you there. This has been a ton of fun. Thank you so much.
49:22 Speaker 00 Thanks for listening. If you haven't already done so, please subscribe to World of DAS wherever you consume your podcasts, YouTube, Apple, Spotify, and more. And please help us get discovered by leaving a review and check out worldofdaas.com. That's worldofdas, D-A-A-S, dot com. And of course, connect with me on Twitter, at Auren, that's A-U-R-E-N. Would absolutely love to hear from you.
Previous Mediaocean CEO Bill Wise - building the operating system for advertising Next Fmr Verisk CEO Scott Stephenson - building a $40 billion data company Page updated November 3, 2026 · transcript and summary are machine-generated