"This feels like 1996": Why a16z's Martin Casado believes the AI boom still has years to run (General Partner)

26:43What is your filter on this market look like? So I've kind of decided I just need to remove... We as investors need to remove ourselves from predicting the future, which is a funny thing because we're supposed to predict the future. I think that's a mistake. And so our approach is very straightforward. We believe that the founder network, the founders themselves are smarter than customers. They see the future, not us. They're definitely smarter than investors. And so if there are three or four very good founders that are working on a space, we just assume that space is good. because A, they're founders, B, they're doing the opportunity cost of doing it.

27:24You know, they're risking their time, you know, their family's wealth in order to do this. And so to first order, we just say, OK, what are interesting spaces? And there's, you know, there's a whole methodology we use to do that. And if there's an interesting space, the next question we ask is who is the leader in that space? And is it too early to determine it? And if, you know, if it's too early, we wait. And if we determine that one that we think is the leader, then we try and make the investment. The thing about this approach is, A, it kind of removes us from, you know, like there's so many aphorisms on investing.

27:58Like this is a great founder and the founder has grit and, you know, like all of these things. But at the end of the day, all of that, you have to kind of filter through yourself and your team. And we're all very biased. And none of it you can systematize. where if you're simply asking the question, A, is this legit space? And B, is this the best company in the space? This is something you could actually throw work on. And it's not, it's clearly not perfect. And in fact, you'll be wrong a lot of the time. But I would submit that if, you know, if you invest in this way, you will be right in a way that's better than market norm.

28:39Do you try, I mean, you must actually, to some extent, still evaluate the founder. And I imagine you've had plenty of meetings where you've, you know, met a founder and felt sort of palpably, uh, this is an extremely impressive person. Do you, it almost sounds like you distrust that emotional response in yourself or how do you sort of think about that? This is a great question. So, so if there's one thing that has shifted in me about how I think about investing and how I think about companies, I used to think from company out. So I'll look at the company. I'm like, the founder is great. The product is great.

29:14The technology is great. The go-to-market is great. I've stopped that. Now I think only from markets in. The reality is the market creates the company in most cases, not the other way around. And so I always start with like, what is the market? And then I ask the question, is this the right founder for this market? The answer to your question of like, is this a great founder or not founder? I don't think that there's a single answer. It strongly, strongly depends on what they're setting out to do. Now, I do weight a lot of things. I do weight things like earned knowledge. Like, have you earned the knowledge to be in this market based on your experiences in the past?

29:51Like, were you at the bowels of Uber building out their storage system and now you're bringing it to the rest of the world? You know, I'm a very product-focused investor. And so I just tend to resonate with product-focused founders that see the world in terms of what is the product we're going to create and how am I going to insert that into the market, as opposed to pure technologists, which don't care about that, and pure salespeople, which also don't care about that. So I'm a very product-focused CEO. But I will say that my umbrella answer, my macro answer to you is almost all questions I ask about companies actually stem from the market on it.

30:26Really interesting. Um, you, you mentioned that you're sort of happy to wait until a leader has emerged in a certain market. Uh, how do you determine when that's the case? And, you know, if it's sufficiently durable, is it like true market share sort of, you know, looking at it from that vantage, or are you sort of making a few guesses of like, you know, maybe. Yeah, yeah, yeah. That's, I mean, that's the, yeah, that's the, that's the part of the job where it's an, it's an underdetermined system, right? There's way more variables than equations and we just do our best. And our analysis is multifarious.

31:02I know as investors and probably fueled by things like X, we like to reduce VC to like, here are these five things. Here's our basic thesis. The reality is most investment decisions take a lot of work. You consider an awful lot of things. And then at the very end, you kind of look at it and you make a judgment on that. So what are the things we look at? Like I mentioned, founder market fit is very important. Tactical approach is very important. The market itself to me is incredibly important. Like I've just learned that if you're selling into a market that's shrinking, life sucks. Even if it's a huge market, if it's a huge, huge market, let's say like switching and routing is this huge market.

31:45But if it's only growing 3 % or it's flat or it's shrinking, you know, you're dealing with budgets that are contracting, people that are losing their jobs. Like all of the incumbents are going to be fighting for their lives. It's just, you know, so I'm very sensitive to markets that are growing versus shrinking. You know, ability to hire, ability to fundraise. I mean, all of these things go. I mean, the final memos for investments tend to be fairly comprehensive. And so all of this also necessarily requires us to do a lot of work before companies are fundraising. And so there's kind of a necessary part of this motion, which is you're constantly trying to enumerate the companies that are out there and then doing the analysis to determine who is in the lead and who is not.

32:30And then you're right. At the end of the day, you just kind of be like, okay, I mean, we did all of this work and we think that you can make this argument here. And we get it wrong a lot, right? There's nobody can predict the future. Yeah, that's the beauty of this asset class, right? Yeah, 100%. I mean, you know, you just have to be comfortable knowing that even if a company looks like the leader now, anything can happen. Like they could get acquired the next day for an acquirer that they decide to do. A new company can show up that didn't exist before. You know, anything could have there could be a platform shift, et cetera.

33:02And so the entire goal is, can you over a set of investments, you know, beat, you know, the upper quartile of the other venture capital firms? That is the goal. And you take the losses along the way. We're talking about the importance of the entrepreneur or the executive. On X, I saw you mentioned that you thought Hawk Tan, the Broadcom CEO, was one of the great CEOs of the past decade plus. And that's not a name that I usually hear discussed in that debate. Can you tell me where that comes from and why you think that? I'll make a stronger form of the same. I think Hawk 10 may be the best, outside of maybe Jensen and a handful of others, he may be the best CEO the industry has ever seen in infrastructure.

33:51He's just unbelievable. You know, somebody should do like the Hawk 10, you know, book or overview or portfolio or, you know, focus piece or whatever. The employee retention is unbelievable. believable. He's managed to do these incredibly complex acquisitions. And I will say, so, you know, normally when you buy a company, any company at all, like the team that you integrate the acquired company into, you know, you've got all these kind of lawyers and corp dev and biz dev and HR people running around. You've got this entire committee for integration. You know, when Hocktan acquires a company, even something like the size of like a VMware, like the M &A committee is Hocktan.

34:37The integration committee is Hocktan. I mean, the guy is just legendary on how hard he works, how he runs his meetings. He knows everything about his business. He knows all of the numbers. And what's interesting, he's a business guy. He's not a technologist nor a product guy. But he has stayed away from the limelight. And to his credit, he just focuses on the business. But there's a lot we can all learn from what he has done and what he's going to do. I really do think he is probably the most iconic CEO right now. Well, you've put a good marker on my editorial calendar there. So I'm going to make sure to do some more research and see if I can write a good story.

35:14I don't know if he's ever done one before, but you should. Yeah, why not? Yeah. That's a great thought. You had another tweet that I thought was really interesting and caused a little bit of a stir in VC world, which it's so fun what things happen to cause a stir or not in these discussions. The tweet, for folks that didn't see it, is the idea that non-consensus investing is where the alpha is, is actually quite dangerous in the early stage. There's a little bit after that, but that's sort of the meat of it. Why do you think that struck such a chord and caused such, not outrage, but discussion?

35:49Well, I think it just managed to piss everybody off. I think there's like every constituency found a reason to hate it, right? The ideal tweet. Yeah, that's right. It's like the mother of all Warshaw tests, right? And, you know, I think for, you know, there's this sense outside of VC that VCs are just pattern matching and add no value. And so for those people, it was a confirmation. And so they're like, oh, I know it. VC is just consensus of us, you know, and now Martine is just acknowledging it, which I totally wasn't. But we can get into that. And then for the investors, it was like an attack on their originality, which was like, I don't do that.

36:27I'm a consensus. You had many junior investors who don't know what they're talking about that they kind of said a bunch of random stuff. But you had some very senior investors that were like, oh, I do all these non-consensus bets and like whatever, whatever. So everybody found like some reason to take umbrage for, by the way, which was like I hadn't even thought deeply about the tree because this is fairly innocuous thing. I thought was just so obvious. I was like, I'll say some obvious thing on a Sunday morning. And it just turns out to have been a lightning rod. What prompted you to say it? And what were you sort of trying to communicate that probably a lot of people maybe talked past the actual point, I think?

37:03Well, I work with a large team of investors, and I'm often in the position of providing guidance. And if you're not considering follow-on capital, then you're not fully evaluating the opportunity set. And I've found that the cliche VC aphorism rule book is like everything must be alpha and this and that. So I just thought there's plenty of people talking about, you know, finding the diamond in the rough. There's plenty of people that are talking about finding the white space. But like there's this another side to it that isn't as represented, which is as you go later and later stages, VCs become more and more consensus driven.

37:47And that's exactly because they're putting more money in and they need more predictability. It follows naturally out of the system. So in a way, this is the most banal tweet you could ever imagine. It's just it's actually totally obvious. I'm not saying I consensus of this. I've done tons of non-consensus stuff. I'm just saying that if you don't consider this, it's dangerous. And so often we don't talk about that. So that was the genesis, which is a totally banal tweet from a very obvious place. Well, it's always good to to cause a little bit of a stir every once in a while, especially over something that is ultimately benign.

38:20I just feel like, I mean, I feel like X is like, it's just totally chaotic, right? There's some tweets I'm like, this is so deep and pithy and nobody ignores the others when it's like this kind of pointless thing. And so in a way, again, just like looking at the market as opposed to the company, I think that tweets are much more indicative of the people receiving it than the person actually tweeting it. Speaking of, well, quite consensus sectors at the moment, let's get into AI and this wild world we're living in at the moment, which you're spending a lot of time on. I know that you have mentioned that some of the energy that you're seeing in AI really reminds you of the 90s dot-com boom.

39:02What are those sort of symbols of that effervescence that you spotted that did bring that to mind? Yeah, so let's see. I turned 20 in 96 and I, you know, I was interning at Livermore, probably starting, I don't remember, it was 97 or 98. But, you know, so I was going back and forth for, you know, a few years. Then I, you know, I worked full time in Livermore in 2000. And I just remember this kind of slow boil that erupted during that time. Like when I started, you know, computer science as an undergrad in 1995, you know, it was kind of this wonky discipline. You know, it was actually kind of in a little bit of a slump.

39:57But the web was just starting and you could feel this excitement. And then by the time I graduated, I mean, I mean, I went to Northern Arizona University. It was a school. My father was a professor in Flagstaff, Arizona. And even in this small mountain town school, we had, you know, students that were graduating, getting these crazy jobs in, you know, as programmers and all over the nation. And, you know, they were being actively recruited, you know, so like there was just kind of all of this excitement. And then when I would go to the Bay Area, you know, I would kind of get kind of caught up in all the founderitis that was going on.

40:35And you had everything at all the parties. You had all, you know, I remember I remember the first time I landed in Silicon Valley. I drove down the 101. I'm like, all these billboards are talking to me. Right. And, you know, there's just this energy and it was in the streets and, you know, you'd have like Linux conference and the Python conference was going on and everybody would show up and all these companies getting created. And it was just optimism and chaos in every sector that you look at. And then it feels to me that things got a bit institutionalized, which is just kind of like another day to do business for the last 20 years.

41:10And I feel, again, now you have a lot of the same type of energy, which is like, I mean, you know, the billboards we've had for a very long time. But again, you've got like these kind of cultural movements that follow it and, you know, all the founders and all the investing going on. I just feel like it has the same level of energy that we had in the late 90s. Do you think we're circa 96 or closer to circa 99, early 2000s? 96. Really? You think we got some room to run? I think people forget what a bubble looks like. I mean, every time valuations go up, people say bubble. I mean, you know, but like, listen, I mean, a bubble is like when you get into like a car and the taxi driver is giving you stock tips.

41:52Like that's a bubble. I mean, remember all of the crazy excesses and, you know, all the crazy blow ups. It's totally, totally different. So, I mean, this feels a lot like early 96. And the big difference is, is then the companies weren't even making money and it lasted so much. By the way, people were decrying bubble in 97 and 98. Yeah, I believe that. And 99 and 2000. Like, I mean, you'll be right eventually. Yeah. Yeah. People were saying it. Right. And they actually had really legitimate concerns. You had WorldCom, which had$40 billion in debt, which is super levered. It was like a single supplier that was underlying all of this stuff.

42:34You could IPO a company with basically no revenue, very little revenue. Many of these companies, these crazy valuations had no money. They were making nothing. And so there was these very legitimate concerns. And none of those really exist today. Right. I mean, you know, the companies that are bankrolling a lot of the infrastructure have hundreds of billions of dollars on the balance sheet. You know, Google, Meta, Microsoft, like OpenAI has real revenue. Cursor has real revenue and and the valuations aren't totally out of whack with with the revenue. So, yes, you know, markets will oscillate for sure.

43:12And so they'll go up and down and you'll have pullbacks or whatever. But I don't think we're anywhere close to like a, you know, late 90s level bubble. No, I think that could come. And, you know, listen, when like, you know. And probably will, right? And it probably will. But like, I don't think we're anywhere close. I just think people forgot what a good bubble looks like. They're a lot of fun, man. So yeah, the fun while the music is on. I promise.

43:49just as fast. Persona helps secure some of the internet's largest and most trusted platforms with identity verification. If you're building a product where trust matters, identity should be a priority. You've probably already experienced Persona without realizing it, verifying your LinkedIn profile, signing up for Etsy, or renting a scooter with Lime. Trusted by leading companies like Square, Brex, and Twilio, Persona gives you the building blocks to create identity flows that adapt to your customers, risk tolerance, and locales you operate in, whether you're verifying age, onboarding businesses, or automating KYC.

44:25It's fully configurable, so you can launch in days, not quarters. Want to see for yourself? Generalist listeners get a free year of the starter plan. Head to withpersona.com slash generalist and check it out. I think, you know, I don't know where we are in the cycle. And, you know, I didn't live through that period as an adult, so I can't I can't compare, but I think we're at the stage of taxi drivers knowing these things very well. I do think the valuations are certainly getting spicy in some levels. Maybe they're not quite at the peak. Yeah, but I mean, honest question for you, do you think right now it's out of whack with 2021?

45:04I don't think it's 2021. Nope, I agree. I think we're not there yet, but I don't know. Does it feel like 2019, mid-2018 to me? Yeah, That seems about right. And so, yeah, maybe we got another 18 months or two years, but I don't know if I'd love like how many if I was writing big checks in, let's say, 2019. I don't know how many of those I would have been thrilled about in 2022. Right. For sure. You have valuations waxing and waning. I think it's great to to actually apply it to 2021. I mean, 2021, there was a lot of excitement, but it wasn't actually driven by real business usage, right? It was like COVID, the flight to online, and then just a bunch of private capital flooded in the market.

45:52Remember, like, you know, Tiger, Co2, Insight, all of these were deploying very heavily. And so in a way, there was kind of this excitement and exuberance, but not for any sustainable business reason. It was really like an influx of capital and then this kind of quirk of the macro that wasn't sustainable. But with AI, I mean, we were three, four years in. It looks sustainable. We understand retention. We understand growth. We understand margins. Yeah. And much less of a tech revelation. Yeah, that's right. So we actually have a foundation underlying it. So I would say, yeah, I mean, it kind of feels a little bit 2019-ish, but it's real.

46:33And so, you know, unlike, you know, the 2021-22 collapse, I mean, you could argue that we're still early in cycle. And yes, it's going to continue to oscillate. But I don't think we're anywhere near the top. Interesting. Yeah, I think that's, I mean, I want to think about it more, but I think you make a lot of very good cases there. We don't have the tigers coming in, but we do have a lot of sort of sovereign wealth fund money perhaps coming in. and a lot of big corporate cash, right? Totally, totally. That's a different level. This is actually very, I mean, maybe on this podcast, we're not going to have the time to dig into it.

47:09This is very interesting construction about the current technology wave is you can actually deploy capital and you can get revenue on the other side of it. And these are very capital intensive businesses, right? And I think that is what the market is trying to normalize. Like you can't even really enter the casino without a billion dollars for these foundation models, for example. And that is because, so I agree we're in a bit of like terra incognito as far as understanding what the capital structure is long time after you've raised this much money. But what we do know is you can actually convert it into revenue and into users.

47:44And so I think this is where we're going to see a lot of rationalization and normalization in the market. But again, I don't think it's basic. It isn't speculative, right? It is just trying to understand what the market is doing. I ultimately think markets are very efficient. And so I think, you know, like we'll rationalize, but there's a true, true value being created in this AI. And I think that if money's not following it, it's going to miss the greatest super cycle in the last 20 years. Yeah, that's the, you know, the other side of it is like you could really miss out. You know, you mentioned that there's really something obviously valuable being created, and I fully agree.

48:22But I was interested in the fact that you see these studies. You know, MIT had their study not long ago that said that, what was it, 95 % of these enterprise deployments are not delivering value. Why is there that gap in what we're seeing? Is that like a measurement problem? Is it a deployment problem? I think one of the problems with AI is that it's been around forever. And so we have all these presuppositions on what it is, right? So here's my view on AI. Right now, AI as it is, is very much a individual prosumer type technology that's attached to individual behavior, right? It's like me using ChatGPT, me using cursor, me using ideogram, right?

49:04Me using mid-journey. And the value that organizations get is that their users are using ChatGPT. Their users are using, you know, whatever. That's what it is. However, there are platform teams within the enterprise and their boards are like, we need more AI, go implement stuff. And so they're scrambling to do these AI projects. And of course, those are failing. This is such a different technology and a different shift. So if you measure some internal effort to go ahead and do stuff by yourself, then I would say the failure rate, of course, is going to be very high. But that has nothing to do with the fact that now many tens of millions of users are using the technologies, technologies, getting value from them and driving that value into whatever the workplace is.

49:54And so I just think that when it comes to this wave of AI, we have to realize it's a very new thing. It's going to have a totally different adoption cycle. We've not yet cracked like the direct sales enterprise. you know i would say for those enterprises that are listening you know rather than doing your own kind of project for now it's probably better to work with like a vendor or a product company that's actually doing these things then over time just like the internet by the way the internet was the same way just like the internet it will make its way into the enterprise yeah in a way that we all understand but but it's not it's just not there yet what are the ways that you've ended up incorporating it into your life most which would you say and and on the other side are there areas in which you're especially protective of not using it sort of to preserve your thinking?

50:45I mean, like I mentioned, so I code with AI. So the reason I stopped coding is I just didn't want to learn the next framework, right? I mean, the thing with developing in the late 90s is you'd sit down to your computer and you'd write code, you know, and it was all kind of there and you didn't have to learn a lot of stuff. You'd mostly just write in code. You know, and then I, you know, through the 2000s, I did my PhD and, you know, so then, you know, I kind of invested enough time to understand all the frameworks and whatever. But I step away because I'm building a business or I'm becoming an investor.

51:15And when I go back to it, I just have to learn all of these new things, especially with all this web stuff. And you're not learning anything fundamental to computer science or anything foundational or anything that's useful outside of that context. You're learning whatever stupid design decision some random person that created the framework did. And so that's really what slowed me down from coding. and with AI, I don't have to deal with any of that. I'm like, whatever, give me boilerplate for an app so I can write a video game and all of those decisions are made by AI. So I use AI coding very heavily.

51:48I do it almost every night and it's really just been lovely. Yeah, yeah. It's kind of my relaxing time but it's really just lovely to be able to just kind of focus on code again. Another kind of just personal thing I like, so I love reading, you know kind of you know historical you know books on historical figures that are closely tied to innovation or economics and often I have a lot of questions and so these days what I'll do is I'll read a chapter and then when I walk my dog this is silly when I walk my dog I use grok audio mode and I actually have conversations about the chapter and in a way I don't even care where if like the questions I have are kind of analysis, synthesis questions, not fact-based questions, like make an argument of why the school of Salamanca in Spain in the 1300s was a progenitor to the Austrian school of economics.

52:44Right. And so like, and you know, so I actually have these conversations about what I read and I just find that I think more deeply about it when I do it. And I actually find it interesting and it's kind of more well-rounded. And so that's personally been great. I think I'm a bit OCD when it comes to writing. So I will not use AI for writing. I think writing is thinking and I use writing to think. And so if something did that for me, I wouldn't be thinking. And I think this for me has just been a lifelong tool. And so I don't think I've ever used AI to write a single, I mean, maybe that's not true.

53:19I don't mean to be too categorical, but I never, never use AI for writing. So that's the one area that I've really tried to protect. Really interesting. I can't help but ask what some of those historical books that you have enjoyed might be. Yeah. So I've just been in Eisenhower lately. I've been going through a bunch of Eisenhower books. It's interesting about Eisenhower, right? He's a conservative president and he was a moderate, but he was also under his watch that the Warren court was created. And the Warren court, you know, I mean, very famously, you know, was the vanguard of the civil rights movement as far as, you know, overturning, you know, policy and kind of getting rid of a lot of the Jim Crow laws, et cetera.

54:00And so a lot of kind of my questions have been, you know, today we, you know, people, you know, criticize the court system and there's a lot of rhetoric on how like the courts being stacked, et cetera. And so I've actually been kind of having conversations with Grok, comparing and contrasting the rhetoric around the Warren court with the rhetoric around the current Supreme Court. It's so interesting how similar the criticisms actually are. And so, of course, it's a different environment in a different era. But for me, I just feel much, much closer to what's going on now as being part of a historical, like a historical trend than some total aberration.

54:34I like to be part of like the broader narrative. perhaps starting in COVID or perhaps you would start earlier even uh it feels very very clear that we're living in history in a way that uh you know maybe wasn't as as obvious you know a couple decades ago or something like that cool is to actually rewind the clock and and and listen to the rhetoric during the Vietnam War and listen to the rhetoric during the dot-com yeah boom and listen to the rhetoric you know around the war in court and realize I don't think it's actually that much, you know, we always have this kind of story. Not so different.

55:07Yeah, we always tell our stories like, oh, this is unprecedented times. We've never done this before and blah, blah, blah, blah, blah. But like they said those words back then too. They did. They were like, oh, this is unprecedented. We've never done this. It's the end of the nation. Like, you know, for me anyways, it's nice to realize that this is a continuum. It's been going on for a long time. The U.S. is anti-fragile. It is the best country, you know, on the planet. You know, like we have always have challenges to deal with and we do a good job of dealing with them. Are there parts of the AI world right now that you consider almost a mirage, you know, something that looks like it could be something, but for some fundamental reason, it's unlikely to last?

55:45You know, I think folks have talked about prompt engineering, for example, is something that's maybe more of a transitory state of affairs. And I wonder, yeah, how you, you know, what you might point to that has a similar quality. So I think that what we're seeing is we're seeing two pretty distinct paths that these AI model companies take. So one of those paths, the model just does more and more and more. And so you basically have one model that does more and more and more. Like these code CLI tools like Codex, for example. So you could be like, I'm going to make it super complicated and do all this prompt engineering and have all this software.

56:26Or I could just expose the model to the user. and it seems in these situations you just expose the model to the user and it just does better because the model is smarter than whatever code you're going to write and then it's just so hard to kind of interpret what the the model is going to say anyway so that's one path it's like you know we're also seeing this in the pixel space which is you know instead of like having a model for image and a model for 3d and a model for music and a model for characters i'm going to have one video model it does everything and oh by the way i'm going to make it interactive this is genie 3 so it just does everything right so there's like one path which is the god model path and the argument for that is the bitter lesson argument that's you know you have all the data you know like the model is smart etc this is kind of you know and that's clearly a viable path the other path is the composition of models paths which is let's take the pixel case again uh i actually i just saw this amazing video um this morning on x where somebody made a video and they're like I used Midjourney to make the images.

57:24I use World Labs for the 3D scenes. I use Suno for the music. And it's this composition of different models. And you look and they have this just beautiful video that was created. And so the argument for the composition is, if you have an opinion on what comes out, you'll just have a lot more control. If I want fine-grained camera movements and field of view. I'll need 3D. Maybe I want very specific images and I want like consistency across those. So I'll need an image model. I want the music a certain way. I may want to change it over time. So I want a separate thing for music. Right. And so I honestly believe we're going to see both of these paths.

58:04And I think the biggest mistake is people assume it's going to be one or the other. Right. Everything's going to be one model. But the problem with that is like composition is just real. We've got an existing, you know, set of tool. We've got existing tool chains that use like components of outputs that you're going to want to use. And so I think that's a mistake. And the other one is like, oh, these big models aren't going to be useful. Like you need a collection of small models. Clearly that's not true because the better lesson will continue to make these single models much more powerful. That's really interesting.

58:34You mentioned Codex there and you've talked about, you know, using a lot of these tools in your evenings, which brings me on to Cursor, which I know you're very involved with. When you were doing the analysis on AI code generation and thinking about, okay, who's the leader here? Was it just blazingly obvious that it was Cursor? How did that sort of come about? Yeah. I mean, listen, it depends on what you're doing. I'm a developer. We were looking at developer tools and the developer tool is the IDE. Now, listen, the coding space is enormous, right? There's repos, there's testing, there's PR management, etc., right?

59:15But in the case of coding, the Copilot had given us a glimmer of how powerful AI could be integrated in the development process. the Cursor team executed so exceptionally well. Half of our companies were using it. And they were just, at the time, just very, very focused on building out the IDE and being the leader. So just for that bet, that was very clear. That didn't mean that we didn't think CLIs were good bets. It was just different, right? That's different. I mean, you could give a... And at the time, there really wasn't as many approaches that were using like the PR for an interface to the developer, right?

59:57Like using GitHub as interface for the developer. But, you know, it was also pretty clear that like Cursor's ambitions was to change all of code. It was also very clear that like code was evolving. And so, you know, from our perspective, you know, a very, very product focused team working on tools for developers that has this kind of broad vision was, you know, the right bet for developer tooling for us. And of course, that's worked out quite well. It's just so important, I said it before, I want to say it again, that that doesn't mean that there isn't tons of value in all of these other areas.

1:00:32Coding models, tons of value, CLI tools, tons of value. I mean, this space is enormous. If you just do rough math, right? Let's say there's 30 million developers. There's more, but let's say it's 30 million. Let's say they get on average 100K a year. I mean, what is that? It's like, what,$30 trillion market or something? Yeah. Yeah. Let's say you get 10%. That's$3 trillion. I mean, we're talking about an infinitely sized market. And if you ask me, what is the one area that AI has surprised you? It's encoding. Listen, I've been developing my whole life and I would never have guessed it'd be this good.

1:01:04And so you've got an infinitely sized market that AI is very effective at going. And so I think we're going to see a bunch of super successful companies. What do you think will dictate the winners and produce real defensibility here? Because obviously, it's given the size of that market, you see lots of interest from large companies and insurgents to sort of take a piece of that space. So my general rule of thumbs is while markets are accelerating in their growth, they will fragment. And that's a natural law of physics. And so everybody worries about defensibility on day zero, which is just dumb, in my opinion.

1:01:45It doesn't matter until markets slow down or they consolidate, right? And that literally just kind of falls out of the, like, listen, if I have to spend a dollar, if I'm in a company and I've got to spend a dollar, why don't you spend a dollar in an area where I don't have competition or I do have competition? Well, of course, you're going to do it where you don't and where you're the leader anyways. And that's why we've seen basically fragmentation in most of these domains. We've got companies growing in most of these domains. And so I think when it comes to code long-term, what keeps them defensible?

1:02:12So here's my current view is I don't think there's any inherent defensibility in AI. I don't think that exists. I think that AI overcomes the bootstrap problem. So it's trying to solve your customer acquisition problem because it's so magic. And that won't be the case forever. But right now, it's like somebody invented cold fusion and people show up for electricity. Right. So like it solves your customer acquisition problem. But from a defensibility standpoint, you have to go to traditional modes. Right. You have like, you know, we know how to do modes. And, you know, whether that's a two sided marketplace, it's an integration mode, it's a workflow mode, like whatever it is.

1:02:45like you still have a company have to build that you know the good news relative to incumbents last last point on this is when you have new behaviors incumbents have a tough time executing and we clearly have new behavior here right it's like it's an individual behavior it's a new relationship there's actually often an emotional component to like you know like the shift between gpt4 and five we saw that and so i think new behaviors advantage challengers i think we're seeing this play out. And so I worry much less about the incumbents. I think, you know, if you're a founder listening to this and you're doing an AI company, priority zero is finding that white space, not worrying about defensibility, in my opinion, you know, and then once you find that white space, you know, rely on traditional emotes to protect it when the market slows down.

1:03:31There's another of your companies that I have been so interested in from, you know, only from the outside that I'd love to hear, you know, your story with them and for folks that maybe haven't come across them yet for them to understand what they're building. And this is World Labs, which, I don't know, reliably anytime you go on X, there is now something really interesting, some sort of 3D model that World Labs is responsible for that is quite - It's magic. It's the most magical thing. Yeah, mesmerizing. So how did that come about? Yeah, this even goes back to my experience with writing 3D engines for video.

1:04:06I mean, It's just of particular interest. So listen, World Labs was created by the true pioneers in 3D. It was Fei-Fei Li, who did ImageNet, super famous Fei-Fei. It's Ben Mildenhall, who created NERF, which is the nearer radiance field. It's Christoph Lassner, who was like Gaussian splats before they were cool. It's Justin Johnson, who was the style transfer guy. I mean, they've just got the most epic team. The easiest way to articulate what they're doing is they want to take a 2D representation, like an image, and create a 3D representation from it, like a scene or a world. And it's a very, very tough problem because if you just have an image, you can't see everything.

1:04:50You can't see in the back of the table. You can't see behind you, et cetera. So there's a ton of generative components to it. And it's also a tough problem because lots of, you know, the way that you train models is with data. And there just isn't a lot of 3D data. So it's just kind of this unsolved problem. But it turns out to be a very horizontal problem, right? Like, so, for example, why would you need a 3D scene? Well, you could use it because you wanted to create a pure virtual environment that you want to interact with, right? You want to place a character with. You want to change the angle with it.

1:05:21You want to augment it. You want to step into it like VR, right? You could also use it for any sort of kind of design, like architecture. You can use it for AR. Actually, I just saw this. This guy named Ian Curtis did this cool thing where he had a 3D representation of his living room in his Oculus. And then he was overlaying changes on it that he made with World Labs. And so he could switch between the virtual labs recreation and the real recreation. So you can like change furniture and change things like that. And then actually, ultimately, this is very relevant in robotics, right? So the problem with just 2D video is you actually don't have depth.

1:06:05You can't see behind things, right? And so you need to create some 3D representation if you want like a traditional program, let's say like a robotics brain, to decide things like, you know, how far away is this? What might it look like on the other side? How do I plan around these things? So the more 3D representation you can create, kind of the smarter, like an embodied AI would be. And so they're really trying to tackle this kind of holy grail of problems of, I just have one view in the world that's 2D, which is what our brain does. And then how do I kind of recreate that in 3D so that I can kind of process it?

1:06:42The robotics piece was, you know, is the piece that I think is so interesting. Obviously, you know, the VR applications feel very obvious in a way, but obviously, on the other hand, still not a market that is massive at this point. But the robotics piece feels like that could be. I don't know, truly a game changer when you're combining some of these other developments we've seen in that industry over the past couple of years. Like it really feels like it's addressing one of the major sort of limiting factors. Well, let's go back to the market size first. It's because I feel this is a mistake we keep making in the AI world, which is nobody would have said 2D images is a market.

1:07:23Nobody, right? There's a whole class of companies which were like, you know, in the past, you know, like they were, you know, small acquisitions. They were never really profitable that we're trying to like, you know, build 2D images. And yet now, you know, we have companies like Midjourney, which famously was bootstrapped to hundreds of millions of ARR. We've got BFL. We've got Ideogram. Very successful companies, et cetera. And so I think in general, when you bring the marginal cost of creation to zero, the market size explodes, right? Marginal cost of image creation, of video creation, of music creation, et cetera.

1:07:54So and again, I know that this wasn't the point of your question, but I think it's very important to touch on is if you bring the marginal cost of 3D content creation to zero, I think that that market is infinitely large. I mean, one of the reasons VR sucks is because there's no content. I mean, I don't know. I've got a Quest 3. I love it. But I go, I spend$24, and I get the stupidest little thing. And so I would say a lot of metaverse, I hate the term, but I mean, just to get us all on the same page, VR, online gaming, et cetera, is really gated on content. It is so hard to build 3D scenes. It is so expensive.

1:08:30And so I think that markets that weren't markets before can become markets. That said, I agree with you that long term, if we're going to have embodied AI, I'm not an AGI guy, but I'm going to say if we're going to have embodied AI, embodied AI that looks at the world and then creates a re-representation of that world and decides how to interact with that world, somewhere, somehow, you're going to need to recreate that world in 3D. Right? You can't do it with length. Yes. Like the description I like to say is like, let's say I put you, I blindfold you and I put you in a room, the lights are off and I try to describe the room so you can navigate it or like, or do any task.

1:09:10Like the words are just not going to be accurate enough. I like to be like, there's a cup in front of you. It's about three feet. You know, like that won't work. On the other hand, if, you know, I give you a camera and then you can kind of recreate the 3D and you're positioned in that 3D, of course, you can now navigate the room. So there's something very fundamental to this solution space for embodied AI. You said a few things there that were super interesting to me. One of them, I do want to dig into the VR piece. It's true that there's not enough content, but isn't the real constraint there the hardware?

1:09:49Like functionally, there's more than enough content for us to live almost infinity lives for us to be in it. But until it actually feels sufficiently high fidelity, it's just not enjoyable enough, right? Maybe for some people. I mean, listen, I love VR. I have, every time I need VR thing comes, I buy it. And my problem is, is unlike a video game, which are deeply immersive and, you know, you've got a ton of content. Like, I walk a plank and then I'm done, right? I shoot a zombie and then I just feel like you don't have enough immersive content. And so there's probably somewhere in the middle. If you look at a lot of online purely virtual experiences, the gating factor is like, how do you build these very, very large worlds?

1:10:34It takes years. It takes teams of people, years, to build kind of these levels and these worlds and these 3D environments. And what's very interesting, I think this is such an important part. It's very interesting. I work very close with World Labs. I go in on Wednesdays. I work with a team. I write code. like you know i mean you know i mean it's all silly you know like i'm like yeah i'm like like a beta user right like i gotta i gotta you know some do some kind of silly things but i'm very very close and they work with a lot of artists and these are traditional true artists that have backgrounds in 3ds and they make these beautiful worlds and whatever they spend a ton of time on they'll spend tens of hours making it and so what you end up with is you end up with a very detailed a very rich virtual world that would have taken maybe a year, you know, if you had a team of humans, they can be like one person can do it with less time, but it still requires a ton of craft and a ton of work from an artist.

1:11:30And so I think that, you know, technology like this is going to increase the amount of virtual scenes and worlds that are there for us to kind of view and explore. And I think as a result, any market that requires these is just going to grow because you can produce more, better quality, and faster. Really interesting. And then you said you're not an AGI guy. Tell us why. I think at the very foundations, I don't think we have figured out how the human brain works. And I don't think, you know, I think maybe a language model or something is a small subset of it. But I tend to agree with Jan LeCun, which is, you know, we'll get to AGI at some point in time and we keep chipping off pieces of it.

1:12:19But like there isn't a straight path from where we are now. It's not like you just add compute and data to the existing models and then we have AGI. I think that we just keep chipping off each pieces of the problem. And so for me, using AGI as some goal or measuring stick or destination, all it does is encourage very sloppy thinking because it ends up becoming the place that you put all of your expectations and all of your fears. And right now it's not even a real place. And so I really try and force people not to use the term AGA, not to trick in terms of it, because it's very hard to have a conversation because it's such a holding place for, you know, magic and magic fears.

1:13:06And so I like to talk about concrete problems, solutions, products, technologies, technology trends, technology directions. And then maybe at some point in time, we will know the architecture that will provide human level intelligence with all the flexibility that can learn just as fast, etc. And then we can start talking about AGI. But until that time, it just erodes conversational quality. It does not enhance it. I fully agree. It feels like it obscures meaning much more than it reveals anything. Yeah, it just doesn't help in a conversation, right? Yes. It really encourages lazy thinking. It quickly becomes almost entirely semantic where you're like, well, actually, what do you mean by AGI?

1:13:50Oh, well, this is what I mean. Okay, well, then this is how we sort of... And also, it becomes the universal justification without having to actually have a justification. Well, why is the marginal risk for AI greater than traditional computer systems? Oh, AGI. That doesn't mean anything. It's not a statement, right? Why is this going to put N people out of a job? Oh, AGI. Now, both of these are great questions. The labor question is an important question. The marginal risk question is an important question. We should have those discussions. We have those discussions not in terms of AGI because that's not a thing.

1:14:27we should do it in terms of like what's actually happening now and in my experience every time you say AGI this is what people use to justify whatever their fear is whatever their concern is or whatever their most optimistic hope is and the problem is when you dig into it it's like this kind of belief that like there's this magic thing that will provide it and so it didn't for me it's conversational and discourse quality that's the problem with the term AGI not the fact that like Someday we will have computers that are smart as humans. Of course we will, but right now that's not helpful. You mentioned that compute and data is not going to be enough for us to have a straight shot to AGI, whatever we might call it, let's say a brain equivalent in every way to a human.

1:15:11How does that impact how you think about the progression of this from an investment perspective? Do you expect continuous large leaps in the capability of these models over the next few years? Do you think we're sort of to expect maybe more incremental improvements from now on? I think we're part of the long march of technology to solving all problems. And even if we stopped AI research right now, there's been enough that's been unlocked to create a tremendous amount of value. And there's going to be new things that are unlocked. And so I just view this as the same continuum that we were on 10 years ago and 20 years ago and 30 years ago.

1:15:50And, you know, we're going to continue to have to unlock new things. And, you know, I just feel because these things are so startlingly impressive that sometimes we kind of don't view this as part of a continuum that has to go. And like we've already solved it. Now we just have to sit back and wait for it to happen. I don't believe that. I believe, like, listen, the way that I view investing now is the same as I did five years ago and 10 years ago. And we need to have more improvements. But what I do acknowledge is that we've unlocked a ton. And so now is a great time to productize and to turn into real businesses, the work that's been done.

1:16:29You've talked before, I think maybe tweeted about the fact that a lot of U.S. companies end up using Chinese open source models. do you think that there is maybe more awareness of why that might not be the best thing uh and uh that that is primed to change or is it something that you're currently quite worried about no i think it's something we should all be concerned with um you know it's kind of funny this is this is the reason why i got so involved in the political discussion which i'll never do again just because it's such a terrible space to be in. But, you know, you had VCs who should know better and who should be pro-innovation talking against open source.

1:17:13Academia was entirely silent. And so it's like the United States just decided that it wasn't going to invest in the number one thing for proliferating technology the way that we see it. And I think largely because of that, like the proliferation of open source has been pretty muted in the United States. And I do think that China really answered the call. They've done a phenomenal job. I would say many of the best AI teams are in China. Their models are many of the best models and they're being used all over the place. And so I think in some ways, we had the wrong approach as a nation and as an industry.

1:17:55Now, that is being rectified. I think that's being understood. But I think now we have a lot of catch-up to do. I think that, you know, like our models aren't the best. And, you know, honestly, a lot of it just comes down to policy questions, right? Like there's a lot of risk to release something open source if somebody else is going to try to find something, you know, copyrighted in it and then sue you for, right? There's a lot. I mean, there's a lot of spurious litigation around these things. And then we have these policy proposals that would be disastrous, like SB 1047 from Scott Wiener. I mean, part of that was actually developer liability, right?

1:18:33So that means that if somebody uses this in a way that caused a mass casualty event, like let's say a car crash, right? You can sue the developers. And so I think from the United States standpoint, we've not done what we've done in the past. We've used the precautionary principle. we've changed the way that we approach technology from a policy standpoint historically and we've done it in a way that slowed down innovation and as a result we're on our back foot and being on our back foot you know with China with respect to technology I don't think is in the national interest and so I'm you listen I've been very encouraged what the current administration has done with regards to AI I think their kind of policy recommendations have been fantastic and so So I am cautiously optimistic that things are changing, but we're not there.

1:19:20We've got a lot of work to do. Amazing. Well, as a few wrap-up questions for you, as we move into our sort of final few minutes here, I always like to ask a few philosophical ones. One for you is, if you had unlimited resources and no operational constraints, what is an experiment you would like to run? Do I have ethical constraints? I'd say no. No, I'd say for the, you know, no people were harmed in the making of this thought experiment. Yeah, 100 % nature versus nurture. I would like go to space. I would clone a whole bunch of people. I would like have a whole bunch of controls. I would like play out their lives.

1:19:57Can I live forever too? Sure. Yeah, why not? No time constraints. Okay, so no ethical constraints, no time constraints, unlimited resources. Yeah, 100 % nature versus nurture. Yeah. And you can imagine how I do it, right? I just clone. I'd have a whole bunch of people. I'd clone a whole bunch of people. and I'd minorly tweak these things and I'd let them live out their entire lives. I'd simulate entire worlds for them and I'd answer the question, what is innate and what is not? And then ultimately, that would be the question on free will too, right? Yes, that's right. You probably have a few things that fall out of that.

1:20:27So that would be quite efficient. You know the title of the, like, you know, after like 300 years of doing this experiment, like the title of the report will be, what does it mean to be human? There you go. Excellent. That's a great answer. What do you think is a tradition or practice from either another culture or time period that you think we should adopt more widely today? Oh, fuck. Siestas. Easy.

1:20:55That's a layup. This is your Spanish heritage, I assume. Yeah. And unfortunately, these days, it seems only southern Spain. I come from the most backwards part of Spain, right? Is that right? Spain and like, you know, like C.S. says, this is a God-given right. I think everybody should take a nap. There you go. Agreed. Final question. If you had the power to assign a book to everyone on earth to read and understand, what would you want to put on their reading list? The weirdest people in the world. That's a good one. You know, David Deutsch's Beginnings of Infinity, of course. Taleb's The Consequences of Fat Tales.

1:21:34I've never even heard of that from him. I mean, it's a, you know, it's a, it's a, it's a tactical, the statistical, you know, the statistical consequence of fat tails. And then Hamming's how to be an engineer. Huh. Could you tell me a little bit about the weirdest people in the world and the final one? I think I know the weirdest people in the world, but there's a, you know, there's a sort of, I don't know how to describe it, a bit of wordplay in that title that reveals something about what it's really about. Yeah. Yeah. I mean, it basically says, listen, the Protestant revolution has changed the way that we associate with ourselves and with each other.

1:22:09And it basically, we used to be very, very tribal. And so that kind of has certain impacts on like trust. And the Protestant revolution kind of forced nuclear families and forced separation. And that required us to be pro-social. And then it also has a second thesis on how free markets also produce pro-social behavior. the reason that i would include it there is listen i think humanity if you just take the long arc here because we're being philosophical like you know like the like the ultimate enemy is entropy it never goes away i don't think any single tribe solves that i think you need pro-social behavior to actually to do planetary level innovation and understanding how we work around trusts and coordination and cooperation is is very critical so listen i don't think it's the ultimate book i think it's great of that i by the way one more book i'd add is is um the end of history and the last man is that what it is i don't know i don't think i've heard of that yeah fukuyama yeah that's a great book of course the end of history yes yeah the end of history i've never read that but i've you know it's phenomenal it's i mean he it was interesting is it historic he's actually recanted on that but it's like this hegelian view of of humans and And his conclusion is like liberal democracy is the end of history.

1:23:27And I think that's being questioned right now. But he does such a great job of taking the Hegelian view that there is this dialectic. There is this evolution of humans. We are continuing to get better. And then listen, he thought maybe we'd arrived. I think the conclusion now is that we haven't arrived. But I love this idea that we as a species are improving how we interact, how we have policies, how we socialize. And so all of these kind of have this general theme of, listen, we as a species are going to continue to solve problems. We're going to continue to have to work together. We're going to continue to have to cooperate.

1:23:59And ultimately, listen, it'll be us versus entropy. No better place than that to end. Thank you so much, Martine. I really, really enjoyed this. That was a lot of fun. Thanks so much. That's it. Thank you for listening to this episode of The Generalist Podcast. Please subscribe on Apple Podcasts, Spotify, or your preferred podcast app. Ratings and reviews help others discover these discussions, so if you enjoyed the conversation, I'd be grateful if you could take a moment to leave one. For all past episodes and more, visit us at thegeneralist.substack.com. See you next time as we continue to explore the future.