Like, what does it mean to be open source for a model? Does that mean open weights, but then you don't have the data pipeline so you can't modify it? You don't have like the underlying code in any scenarios, right? So yeah, so you don't have underlying code, but you also like it doesn't really matter if you hide it behind the API. It turns out you can distill it anyways. This to me is like the paradox of this AI stuff is the more of an opinion and taste you have, the harder time you're going to have using it.
Hello, fellow data nerds. My guest today is Martin Casado. Martin 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. Martin, welcome to World of D.
Very happy to be here. Super excited.
Now, DeepSeek has been out for now a few months. And now that the dust has settled, like what's the significance of the release?
So there's kind of two theories on this. And so 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 almost like 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, you know, with a lot of fanfare, and they've got, you know, high valuations to hit. We talk about how tough it is in AGI and this and that and the other thing. And it just turns out, especially with these reasoning models, they're not that big of a step from, you know, especially once you see what other people have already, right? So I've got more like it's actually very hard to have gotten to the first breakthrough, which is all the pre-training stuff. Like to get to GPT-3.5 is tough. And you know, DeepSeek did that with V3, and we don't know how much they spent to do that, but like 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 kind of mystique around these things that may not be warranted, and DeepSeek kind of shows that this isn't that hard.
So how do you feel like there's at least 10 companies probably spending like a billion dollars or more on like these models? Like how does that play out and what's like the second-order effects of that?
Well, this is the big question, right? And so I think this is basically what's consensus, which is the first wave was getting to, you know, the multimodal model via pretty aggressive pre-training, right? And this is GPT-3.5, GPT-4, you know, Llama, Anthropic, and this is before the reasoning models, and that was a very, very real breakthrough. And what's amazing about those models, they're incredibly general, like you can apply to the, you know, so you kind of consume all the tokens in the world, you can apply them to all, you know, a bunch of problems, but you know, because there's a finite set of tokens in the world, they all kind of converged to the same capabilities. So now there's a big question is okay, so but do seem like slightly different, like maybe Anthropic is a little bit better for software development and...
No, no, this is the big question, right? So this is exactly right. And 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 or special collected data in order to make it good at certain domains, right? Like you can make it good at 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 kind of 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, like are they going to fragment because the RL stuff doesn't generalize or do they not fragment? I think I'm in the camp of they will fragment.
I think this is what... and mainly you think they'll fragment just because it's a better business model because like 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.
That's a great question. No, I think they technically fragment. Like if you make a model very good at code, it's probably not going to be as good at something else, right? And so if you focus on a certain domain, then technically that doesn't generalize. You ask another amazing question, which is like one of the stories of AI has been fragmentation. I mean, if you think about on a business model standpoint, not in a technical standpoint. So for example, you know, 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 you know, it seems to have not become the leader in high-quality video. It was of course the first ChatGPT. It is the leader in ChatGPT. So the markets are so big and they're growing so fast that independent of the technology, it does seem that they are fragmenting so that you've got leaders emerging in all of these other categories that are showing up. So I think that it's pretty reasonable to bet because the markets are so large and they're growing so fast that we're going to continue to see fragmentation, like a bunch of different companies in different spaces.
And that's because of the underlying data or is that just because like they're just going to keep focusing on and they're going to get feedback from the customer?
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, you know, 3 to 5x or something crazy like that, why would you even do something else? It just doesn't make sense from an investment standpoint. And I mean, even think about just image alone, right? So before AI, you know, 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, right? Like Midjourney, Ideogram, BFL, and now we have multiple and they're at scale and they all are kind of going after different areas, right? Like Midjourney is kind of like gritty sci-fi, like, you know, Ideogram is 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 fragmenting, 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 because it's like you're getting data, you're reinforcing it, and it seems like everyone is super competitive.
Well, I don't know. I do think that we're seeing this break apart. I do think the winners are like in the lead. You just have to be at the right zoom level to see that, right? Like OpenAI definitely is dominant because of the brand too, right? Not just because of the product.
No, this is a very... So, you know, listen, Midjourney is clearly the leader in text-to-image. Midjourney is kind of hard to use. Like you have to go into Discord and then you like, and then later on somewhere in the Discord chat you see your thing and it's kind of annoying. And yet, and yet they bootstrapped to actually quite a bit of scale. Totally. I would say Anthropic is the leader in code. I would say it's probably got 80% of the market. I would say ElevenLabs is the leader in text-to-speech. There's nobody even close to ElevenLabs for that. You know, and so 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 again, I do think basically marginal cost to do stuff has gone down, markets are expanding very quickly, as a result you get natural market fragmentation. I mean, just one more very quick example on this, if you look at the video companies, they even fragmented, right? It was like Pika kind of went consumer, you know, really good at anime, like Luma, you know, focused on creatives, Runway on creative professionals. And so you're seeing this kind of fragmentation all over wherever you look.
And you mentioned Cursor. Like, do you think that they've been so successful and so dominant because the product is so much better or is it just like one of those things where everyone starts using it and then everyone tells everyone else to start using it and then, you know, etc.?
So I think in the case of Cursor, it's because Microsoft had matured the market for two years, right? So Microsoft Copilot came out almost exactly three years ago, and so it had been out for at least a year by the time Cursor came. And Cursor, people are used to using that and they're used to using VS Code, you know, and then they kind of used that as a starting point. And then from there, I think they've executed just exceptionally well. I mean, they just, you know, and so I think it's the combination of the two. I mean, I would say, you know, almost 80% of software developers I know use Cursor. Like it's just insane. Like, and I never even heard of the company like a year ago or so.
I know. I mean, we did a survey of a number of portfolio companies, more than 50% of the developers use Cursor, and like that's against VS Code which has been around for a very long time from Microsoft and also all these other weird things like engineers love their weird IDEs, but they're switching over for a reason, right?
Yeah, yeah, yeah. No, it's been very dramatic to see that.
And what about like there's these like front-end versions of Cursor? There's like the, you know, the Bolt News, the Lovables, the Versels. Those seem to be more competitive. They're all growing super fast.
Of course, they're all growing super fast. So I think there's kind of two markets 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, right? Like the developer understands the PR, understands the code, and that's really squarely what Cursor is going after.
Yeah, as far as they're going after a real developer, not just like a hacker, right?
Well, 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. Probably they have a CS degree or something, you know. They probably built a lot of systems, etc. And it's interesting, like one of these kind of high-level observations I've made talking to a number of users of Cursor is like they're starting to say like, listen, it's actually easier for me to use Cursor than work with a junior developer, right? And so like this is like really these kind of 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, right? And so, you know, like the interface to them is more the final product and it's not the actual code itself. And I found a lot of their users are like the product manager or something, which is like a, you know, fairly technical but doesn't have a computer science degree.
That's exactly right. And so I just feel like that's almost like a different... Yeah, like a different market, a different use case. And at least when I've looked at like the Lovables, Bolt News, Versels, I think they have a higher churn because like you're using it for like a specific project and then you might stop for a while, whereas like I imagine a Cursor IDE, if you're in it like every day forever, like you're not going to churn unless you just move to another product or something.
Yeah. I mean, I think you can almost kind of liken the, you know, prompt-to-full-website 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.
You know, it's like it's almost like this is like the opinion-free version, right? You get a few things to select and they look marvelous and they're dazzling, but like you don't have the prerogative as a user to specify, you know, exactly what you want, right? And so it's good for a certain type of things, but not when you're the super professional, like you actually care about everything. Where on the case of Cursor, you know, this is for, you know, 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. Because 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 that they're non-ambiguous. They allow you to describe things specifically. And so 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 that are like, you know, they don't have, you know, they just want something. And I think that's a different market.
When I look at the developers using things like Cursor, yeah, it seems like the person, like when I see the ones that are getting more benefit from it, it seems like there is a difference in the personality in them where they're willing to like go back and forth in a weird English text way saying lots of pleases and thank yous. And 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?
Yeah, I mean, it's so funny, like there's actually all sorts of personalities for interacting with these AIs, right? Some of them are like what you're saying, which is like you're kind of 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, like literally say like, if you don't do this right, I will kill 10 kitties. And so it's like this kind of very kind of aggressive, like putting a lot of pressure on that. And honestly, there are others, and I kind of fall in this camp. So I mean, I program with Cursor, you know, most evenings.
If you're enjoying this podcast, please sign up for our newsletter. Go to worldofd.com. The link is in the description below. Now back to the video. So 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. And that way because I feel like I...
Is that just because you're more old school? Like if you were 20 years younger, it might be different. It's a very reasonable question. I mean, it could be this is my boomer moment and I should just kind of let go and like let her write a bunch of code. Like I think it's a very reasonable question, but for me, like 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 like we're almost at like we're like going to build in the end of opinion. Like systems builders going to stop having opinions. They're like, you know, not only like do you build it for me, but like you tell me what I like, just because having an opinion on what these systems look like 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 like if I think of like the software developers, they're literally in the tools every day. When I, if I write a piece or something, like I'm using it as like an editor, but it's kind of like at the end and or maybe every once in a while to like help me rephrase something, but it's not like literally in the system as I'm writing. And I would think that would be so good. I just don't, I haven't found the tool to do that. Is there something like that that you've seen?
I really think it comes down to like the user's need for specificity. Like I think you probably have independent taste and an independent aesthetic and an independent idea of what it should come out with. And if you have that, there's no way a model is going to be able to do that because, you know, they produce, you know, whatever they decide to produce. Where I think other people, they just want something and if it's good enough, it's, you know, like they just don't really have an opinion on specifically what that is. And I do actually find like a lot of the uses break down to this. Like I'm the same as you. I use AI all the time, but I very rarely have it produce stuff that I use, you know, 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.
This is when I go and 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...
Because you're not an artist, right? Like so like you don't have the same aesthetic or the same taste. And so I do find the same thing between junior and senior programmers. Senior programmers when they use AI, they're like, okay, write this test for me, you know, help refactor this thing. But they're kind of...
Tests are, you would think would be great for tests and stuff like that.
All the senior guys hate writing tests anyway.
So it's kind of what you'd kind of like, but there's no way they're going to kind of give up the system design.
Where, you know, 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.
Where do you like, there's just been like this open source versus closed source debate in the models, and like where do you kind of come down on that?
I don't know. I don't think there's any parallels to traditional open source. I just don't think there are, right? Like, what does it mean to be open source for a model? Does that mean open weights, but then you don't have the data pipeline so you can't modify it? And then even if you don't have the open source, you can still distill it by, you know, by asking a bunch of questions. By asking a bunch of questions. So I just think it's actually sufficiently different because you don't have like the underlying code in any scenarios, right? So yeah. So you don't have underlying code in any scenario, but you also like it doesn't really matter if you hide it behind that API. It turns out you can distill it anyway. It almost feels like it's just much less of an issue than we've seen in actual open source code, which is like 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. I mean, I think that's kind of what it comes down to. So I just think like the gap between like quote-unquote open source and not is just much, much closer in the AI wave. And so I think we made it a kind of a bunch of to-do about nothing. I think like if companies want open source, great. If they don't, it's fine. And I just don't think that that's a, you know, like there's a major difference.
Until like about a year ago, I don't think I would have ever invested in a SaaS company without like super technical co-founders. And I would say recently I've seen these SaaS companies that like have built pretty cool things where the founders really weren't that technical. They were like good product people and they had really good taste, and they were able to make these products with like the current tools. Is that like changing the way you guys are thinking about investing in people?
Yeah, this is a funny, this is kind of a loaded question for me. Because I tend to historically not think traditional vertical SaaS founders were technical. I've always found that's true. Maybe they weren't even that, that has to be all CRUD. You know, you can hire a random undergrad to build it. There's like, they don't differentiate on technology, etc. And like 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.
Okay. Yeah, that's fair. Yeah, I do think like it's like that product person with taste is just becoming, they're always important, but they're becoming like even more important. You know, and it's very hard to assess that person based on their resume and stuff, whereas like a technical person you can much better assess based on their like LinkedIn profile.
Yeah, for sure, for sure. But I also think that a lot of it comes down to business model too, right? And so like, I mean, like if it's like a product-led thing, then you know, you want to, they have to be really good at brand and, you know, growth features in the product, etc. That's very tough to assess. But a lot of vertical SaaS is like, you know, here's this market and we're going to write some software for the market and here's why, like, you know, we differentiate on a business model standpoint or there's lock-in or some other type of thing that goes in there. Yeah. Yeah. But to your point, I just want to say like, this to me is like the paradox of this AI stuff is the more of an opinion and taste you have, the harder time you're going to have using it because these things are so unwieldy, right? Like they don't, like it can't read your mind and you can't actually describe taste very well in English. And so the more opinion you have on the output, like the more disappointed you're going to be. And so one of two things happens is like, well, the professionals tend to use AI for the things that they're not professional at.
I actually just had a great tweet about this. It's basically it'll help you do the stuff that you're not good at, but like me with images, like I was never good at that and now I'm like decent just because I could write some prompts.
Yeah, exactly. Right. So it's like, 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, right? 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. 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.
Yeah. What there is just the UI of the prompt is weird. It's kind of like, you know, there are not that many people that like using the terminal on their desktop or something, right? It's just like, it's just an odd prompt to go through and I imagine it's just like only a very certain type of personality of a person is going to like the prompt. And it seems like all the UIs today I've seen from AI companies like all start with the prompt. But maybe only like 5% of people are really going to thrive with the prompt. Are you seeing like other types of UIs that are coming out?
I mean, I think the UIs follow traditional modalities, right? 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. 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 like compose, you know, some complex, you know, creative thing, you use something like Comfy. And so I actually think if you actually drill down into what UIs are actually used, they are converging on the traditional UIs. Like if you look at Ideogram, like it starts to look like Canva. If you look at Runway, it starts to look like these video editing tools. And Comfy UI, like it looks like these kind of like big, you know, 747 cockpit, you know, complicated workflows that you see in, you know, traditional like VFX environments or, you know, creative environments. And, you know, I think 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. And so you have these companies that start with like there's a very simple interface to model like text to like whatever the output is, but the companies that do well actually do build out these kind of longer products. And so I mean I agree the text-to-whatever is incredibly basic and not long-standing, but we're already seeing how that then becomes, you know, what we would consider pretty traditional interfaces.
And are there like, I mean in some ways the most traditional interface is I'm emailing or calling or texting a coworker and asking him to do something with me, right? It's like in some ways like that's kind of like the one we're all used to. Or I'm slacking, you know, Martin and I are working together and I'm going to slack, I'm going to Martin at 816Z whatever and send it over to Mike. Like can you just imagine that's the interface?
Listen, I think this is kind of been one of these kind of bad analogy spaces where people are like, why don't like we all should use audio or we should all use text, but we have to remember that before we had computers it was all audio and we decided to build different UIs. Like that was like that was the standard, right? And before, you know, we had complex, you know, 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. So like let's take like this story, like everybody, you know, I very commonly hear, well like this is all going to be audio. And to me, like listen, 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. Like I want to see like, you know, it's not good for me to provide spatial input to the AI. Like sometimes I want to provide spatial input to the AI. And so I actually think the, also more people, some people are more auditory, some people are more visual, like everyone is different, right? Yeah, yeah. And like listen, we have, so let's talk non-AI, we have audiobooks for like that experience, we have textual books for that experience, we have UIs for that experience, and we use all of these things. It's not like one has ever replaced the other. And so I think that with AI it'll be the same. Sometimes it'll be audio, sometimes it'll be text, and like there's a reason we've had multiple modalities of UIs.
Now you mentioned like you just mentioned earlier you're coding like in Cursor every night. Like what are you doing? Like what are you writing?
Honestly like my relaxing time is writing kind of retro video games with Cursor in Java.
Okay. Cool. Just like for fun.
It's okay. It's not like you're like building tools to make your investing better or something like that.
No, no, no. It's just honestly for fun. I mean, but also I mean I, well, I will say that like I do get a good sense of the tool chains by doing it, right? What's useful, what's not. I'm on the Discords, you know, of the other people that I wish this worked better, you know, bugs. Yeah. But also I, you know, I'm part of the communities and the communities are talking, so I get a sense. So I think it does help a lot for because I'm an infrastructure investor, like I invest in dev tools. You got to be using the tools otherwise you'll atrophy, I assume, right? And to be part of the communities because there is a zeitgeist and the zeitgeist, you know, you can only, you can't like 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 this, I mean, I've been in computers since the 90s, right? This is 30 years, right? And so it's just kind of a first love of mine.
And like are there things that you use that you're using AI for as a venture capitalist to be a better venture capitalist?
Well, let me just split apart that question. So I am not convinced these models are good at anything predictive.
Yep. Does that make sense? Of course. Yeah, because it's predicting the past.
It's very good at predicting the past, right? But I think a lot of times people are like, oh, like these things will like, you know, tell you, you know, like they will give you like guidance what to do in the future and predictive. 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, right? Like and I'm learning these days. Actually, maybe one of the biggest personal changes is like I talk to Grok a lot rather than reading books. So I'll have a book that I'm reading that's in like 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, you know, twice as much time just having a conversation with Grok to understand it deeply, like cross-referencing it.
Why do you like, it's interesting, I do that with OpenAI, and I'll go for a walk sometimes in a conversation like this with like somebody like you who's much smarter than me and I don't want to admit that I don't really know it, and then I'll go for a walk and I'll just like ask like OpenAI like, hey, like they were talking about this thing, and then I can kind of get it at my like exactly at my level, like not below, not above. But why do you use Grok? Like what's, is it better for that?
These products are evolving and moving so fast that like my opinion on a product could change two weeks later.
Totally. Yeah. And like by the time this podcast comes out, 100%.
And I couldn't be back. So listen, 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 had a very fluid conversational experience. So what I like to do is like read a chapter of something I find interesting and then like I like then I like to like basically use Grok for steelman arguments. Like 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? What influenced this type of thing? You know, if I were to like, you know, construct an argument of why this doesn't make sense, how would I do it? I mean, it's very good for steelmanning. I find that like kind of just helps my cognitive understanding, my critical understanding of any given space. And it's just again like the product experience for that one week when I made the decision that was better is probably totally irrelevant now.
Are you, I mean, are you, I mean obviously like I'm assuming you have like an AI notetaker on the call and stuff like that, but are you doing other types of things to like help you assess or, you know, I can imagine like an AI founder personality check or something like that.
I don't do anything, I don't do anything predictive with AI. I mean, I code a lot enough with these models. I mean, I spend hours a day with these models that I, you know, it's very clear that they're giving you some, you know, stochastic smoothing of the training set, 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 like, yeah, like to take notes.
I mean, I imagine one of the biggest time things on your, if I looked at your calendar, would be meeting with companies, taking the first meeting with companies and stuff like that. Like, could you imagine a Martin agent talking to the founder agent for the first call or something?
Or I don't think so. I mean, not with the technology that we have right now. I do feel, so let me maybe just say what kind of our basic investment process. Our basic investment process is we kind of carve up the market into different spaces, right? And 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, right? Because founders are smarter than VCs, certainly smarter than me, right? So we're like, okay, there's three founders in this space, it's probably a good space. Then our entire job is to understand between those three companies, like which one is the best. And in my experience in 10 years of investing, well over a hundred deals, is the most important thing is kind of the first derivative, which is like how did the companies fare over knowing them for years or months, like how over time do they execute? And it's just you just need to have multiple interactions to do that.
And so like maybe like AI is one way just looking at like the product velocity then and things like that.
I think everything. I think it's market velocity, 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.
You're all in on, you met with everybody.
