Chris Dixon 谈如何构建网络、运动与 AI 原生产品

Whether you're an investor or entrepreneur, the most important thing to start with is to look for these forces—to look for these exponential forces. You can do all sorts of tactical product things, everything else, but these forces are going to overwhelm you, for better or worse.

How intentional do you think you have to be as a founder about building? You're building a tool—do you have to be thinking about the network as a priority, or can the network sort of emerge? Because in AI so far, we've seen a lot of tools and not a lot of networks. What's your instinct?

I'm super excited and honored to have my partner, Chris Dixon, here today. Chris, you're probably best known for your work in web3 and network effects recently, but what folks may not know is that you led a lot of the most important consumer investments at Andreessen Horowitz, and prior to that, you also founded 2 consumer companies. I thought a fun place to start would be networks. That feels like the first place you really cut your teeth. So maybe talk about your investments in Stack Overflow, Pinterest, and Instagram, and how you generally think about consumer networks.

So many of the most important internet services are networks, right? Going back to the early internet, email and the World Wide Web—which are still, of course, around and really important—are networks. They're networks in the sense that the service gets more valuable as more people use the network. If you were the only one on email, it wouldn't be particularly valuable.

During the rise of the internet in the 1990s and 2000s, that's when you had things like YouTube and Facebook, and later on Instagram, along with a whole bunch of other really important networks. If you're an entrepreneur or an investor during that period, they tend to be very valuable companies. They're very hard to build, and we can talk about that later. There are different tactics and strategies for doing that.

My background is that I started 2 companies. The first was a consumer security company, and the second was a consumer AI company. Then I was a personal investor. I co-founded a seed fund called Founder Collective, which was an investor in things like Uber, Venmo, and Stack Overflow, as you mentioned. I was involved in a bunch of these networks as the internet evolved.

I think to really talk about networks, though, it's important to step back. To me, a foundational question in tech is: Why do these companies come out of nowhere and end up being very impactful, having hundreds of millions or billions of users, and being very valuable in a way that you typically don't see in other industries? What's fundamentally different about tech?

I think the answer is that in tech, you have some very strong, exponential, superlinear forces. The most famous example of that is Moore's law. Moore's law is the idea that roughly every 2 years, or 18 months, the performance of semiconductors doubles. It's a rough approximation, but it's basically been true. You've seen this compounding improvement in processor performance.

I think there's also a broader Moore's law, which is that storage, networking, and all kinds of computing resources have gotten much better. That's why you have things like mobile phones. If you go back and look pre-iPhone, mobile phones were pretty junky and limited in capability. They didn't have touchscreens and had poor performance.

What Steve Jobs and Apple saw brilliantly was that, actually, by the way, the first iPhone was also quite limited. I was one of the people who bought it on the first day, and it was quite limited. But part of their brilliance was that they saw this curve. They saw this exponential curve and rode it. Moore's law is a very important exponential curve.

The other 2 really important exponential curves in software are, first, what I call composability. Composability is really what has accounted for the rise of open-source software.

Why did Linux go from a hobby project in the 1990s to the dominant operating system in the world today? A lot of the answer is composability. Composability means the software is open source, so anyone can contribute to it. More importantly, you can harness the collective intelligence of the internet, as opposed to locking it up and relying only on your employees.

Anyone in the world can contribute. The famous phrase is, “All bugs are shallow with enough eyeballs.” Really importantly, with open source, software becomes like Lego bricks, where anyone can take a piece and reuse it. You get this compounding, exponential improvement and growth.

The third really important exponential force in tech is network effects, as we were talking about. That's why networks are so important. They often start off quite limited. Facebook was just at Harvard, and it was essentially a real-time yearbook for students at 1 school. Then, of course, it hopped by lily pads to other schools and high schools, and eventually to the global domination we have today.

Mark Zuckerberg and the team saw the power of network effects and rode those network effects. That's why Clay Christensen calls this “disruptive technologies.” It's this puzzle of why, in tech, you have these very strong incumbents who seem to miss the next thing.

You could tell the story today about maybe Intel and Nvidia, or even ChatGPT and Google.

Or even ChatGPT and Google. I was just reading about that.

That's a great example. Neural networks 10 years ago were kind of toys, right?

Yes.

They were cool, and a bunch of people saw their potential, but the reality is they just didn't work that well. I remember there was a chatbot—I want to say around 2016 or something. If you remember that, chatbots had a moment back then.

They had a moment, but the reality is they weren't that good. They just couldn't do the job. But, of course, they got much better, and the genius of OpenAI and other pioneers in the space was to make that bet, right?

That's right.

Google today is in an awkward position because it has this huge incumbent business that depends on sponsored links, and it's trying to layer in AI and do things like that. In some ways, it didn't come out of nowhere, but it grew faster than even some of the optimists predicted. It improved faster.

The big takeaway here is that whether you're an investor or entrepreneur, the most important thing to start with is to look for these forces—to look for these exponential forces. One of the lessons I learned in my career was that you can do all sorts of tactical product things, but these forces are going to overwhelm you, for better or worse. The first thing to understand is the landscape of these forces, how they're moving, and how you can hopefully be on the right side of them.

How intentional do you think you have to be as a founder about building? You're building a tool—do you have to be thinking about the network as a priority, or can the network sort of emerge? Because in AI so far, we've seen a lot of tools and not a lot of networks. What's your instinct? And, of course, in hindsight, everybody was designing a network from day 0.

That's a great question. I wrote a blog post years ago called “Come for the Tool, Stay for the Network,” and the idea was that I was observing what I saw as a tactical pattern among entrepreneurs. I cited Instagram as an example. Young people won't remember this, but Instagram's network initially was not a big part of the product. It had a button where you could share on Instagram, but why would you do that? No one was on it.

What you would do, I think, was 2 things. First, Instagram had these cool filters, which at the time you had to pay for on other services, and they gave them away for free—effects or lenses, whatever you want to call them. Secondly, they piggybacked off other networks. You'd share to Twitter. Then, I think a year or 2 later, Twitter blocked them, and there was a whole thing.

You see that today, maybe with Substack. Substack starts off piggybacking on the email network and on Twitter. My sense is that they're now getting traction with their own network. You go to the Substack app, right?

That's right.

I think it's a similar tactic. You can see some of this “come for the tool, stay for the network” pattern in modern productivity tools. I'll defer to you on this because I'm not as up to date, but

Think about Figma and Notion—things like this.

They're useful as single-player tools, right? You can just go to Notion, and it's a really nice way to edit a document, or use Figma to do design.

But there are also social features that I think become essential.

These things are always a matter of degree. Google Docs—I love Google Docs, and I use the social features. The reality is, is it really a network? I could probably switch and just share links with somebody else.

But the social features layer on. For some products, like Instagram, they become essential. You simply can't leave Instagram if you have a following and want to keep that following.

So it kind of varies by use case. But I think the point, by the way, is that you see some of this now in Stripe doing the Link product, which is a payment app. I think Shopify and the Shop product are a really nice user experience: I don't have to type in my credit card again. Now there's kind of a network, right? Shopify originally was just kind of a tool for sellers.

For a merchant to get online.

That's right. So I think it's a really powerful tactic, right? Because network effects cut both ways. Network effects are great when you have them, but they're really hard in the beginning. No one wants to be on a dating site with 2 people, right? How do you make these things useful from day 1?

The problem with single-player products is that they're just hard to defend, right? I think you're seeing this in AI today. You're seeing a lot of really cool tools because it's an amazing technology, but then it's like, okay, you can change your face with FaceApp or whatever. But how does it move beyond faddishness? How does it move to something that really engages people over a long period of time?

Often, the answer in consumer products is networks. So then you have to layer in a network. The challenge, of course, is that you don't want to just layer it in for the sake of it. You need it to actually be useful. So, yeah, I'd love to hear from you. What are you seeing in that area?

Yeah. Well, it's actually interesting. It feels like the big networks have become hypersensitized to this idea of new networks emerging that were bootstrapped on their networks. I think Twitter 10 years ago would have been a lot more asleep at the wheel to the threat of Substack, and they were pretty aware of this potentially happening. Of course, Facebook has deplatformed a ton of companies that they thought were going to do this, as have Instagram and others.

One, I think the networks are more sensitive. Then, on the tool side, the tools have been specializing in their own directions. Part of it is product features, but part of it, even for some of the multimodal tools, is aesthetics. Midjourney just has a different aesthetic than Ideogram, so they can both coexist and they're not directly competing. Even though the tools are seemingly substitutes, so far we haven't seen that trade-off, and they're all working. Maybe that's just where we are in the product cycle.

But I do think it's a topic for a lot of AI founders: there's not an obvious network to build around a lot of these tools. How much of that should be predesigned versus, “Let's just keep pushing the edge and the network will emerge”? It will show up in 2 ways. One would be in usage—you might see some of these tools not get used as much—but the other is in pricing, right? Like—

Yeah—

Even if you carve out a niche, how much more are people willing to pay for that niche versus those competitors, right? So—

Yes.

Yeah. And actually, prices have been going up, interestingly. Google's top SKU is $250 a month. Grok is $300 a month. I don't think we've ever seen a time when consumers were paying those kinds of prices. One of our extreme views here is that the future of consumer disposable income will be food, rent, and software. Software is going to subsume a lot of the other areas of discretionary spending today.

Yeah, it's also possible. I've always suspected that in tech, we in Silicon Valley underestimate the power of brands and consumer inertia. I think you're seeing that today with ChatGPT, which became such a household name almost overnight.

Even though it doesn't have, in the technical sense, maybe network effects—I mean, memory and things, but that's more stickiness than network effects—the brand effects are so powerful, right? Cursor is known as the best vibe-coding platform or whatever.

That's right. Yeah, I was going to ask you about that, Chris. Of course, network effects are the gold standard for defensibility. You've maybe talked a little bit about how brand is underappreciated; you just mentioned it. Do you think being a high-NPS, DAU product is enough of a moat? Or do you think that we really have to push for building around these compounding forces?

Yeah, it's a really interesting question. One argument would be that the internet—I think there's a decent argument; I was actually having this discussion at one of our partner off-sites—is that maybe a lot of the network effect has been externalized to the internet. The idea being, you have Cursor, and then suddenly it becomes popular.

Or Midjourney, let's say, and then you get all of these Midjourney influencers, YouTube videos, and how-to guides. So you still, in some sense, have a network effect, but it's not a network effect that's in the product itself. It's sort of externalized to the internet, right?

You try to show up top in search; ChatGPT recommends you; the algorithms feature you. And, of course, there's a soft sense of a brand—people have heard of you.

But it's also this whole giant kind of system, right, with all of these different interconnecting networks that might strongly favor those products. Then it becomes sort of a timing thing, right? You get in early. The timing seems quite important: being the first to own the meme in the category and get that effect going, and then maintaining it through product velocity, high quality, and everything else, which is nontrivial.

It's very hard to do, I think, particularly in AI. You tell me, but

To always stay on the cutting edge is expensive and takes a lot of capital. That's another thing, by the way: the capital effects in AI.

Right. You do well, you raise the most money.

I assume the people raising $1 billion have already proven a bunch of things, and at some point the capital becomes a moat, right?

100%. Yeah. No, it's very interesting, because there's this barbell that's happening even in software, where the bigs are getting bigger, but we're also seeing the single-person, $100 million run-rate company coming—or maybe it's already here. Certainly, the bigs are getting bigger, and capital is a part of that.

And maybe the market's just so big that the answer is both—all of the above. It may just be that, as you said, it becomes like food and rent, and software is moving beyond the quote-unquote software budget.

It really hasn't been zero-sum so far. It's been shocking: prices are going up and everything feels like it's working. Maybe we'll look back and say that was a sign, but so far, so good.

You know, Chris, I thought, actually, since you mentioned vibe coding, it would be fun to talk about movements. It feels like you've been early to a bunch of movements. Products like Coinbase, of course, and MakerBot—those felt like niche communities on the internet when you started paying attention to them.

How do you think about investing in movements, and how do you think about building around them when there are questions around, “Is this a toy? Is this something structural? Is it durable or ephemeral?” Maybe talk a bit about that.

Yeah. I mean, it's a little bit to the point we were talking about, about the networks becoming externalized. I used to spend a lot of time, 10 to 15 years ago, on subreddits and niche communities, partly because I'm interested in that stuff and partly because I think they're very powerful.

If you look at Wikipedia, Stack Overflow, and a lot of these interesting movements—community sites—they're often 20,000 people. They aren't that many; they aren't the millions that you might think. There are millions maybe doing a little bit here and there, but I just think a lot of the things that have been popular movements that grew were really led by a relatively small—I mean, on the internet scale, relatively small—hardcore group of enthusiasts who are really smart and often technical.

There's this famous old quote, I think from William Gibson, that “the future is already here; it's just not evenly distributed.” I've always believed that. If you just go back historically, that's the case. We're talking about neural networks; that's been going on since 1943 or something. There have been communities of people, including a lot of the people who lead the labs today, who 15 years ago were seen as niche, or more niche, or something. Neural networks weren't the dominant approach.

With that thesis, you want to find the next thing, the next big thing.

One way to do it is to look around and see where these hyper-enthusiastic, sometimes cultish communities are. They have their own language, their own norms, and a sense of insiders and outsiders. I got into that kind of thing a while ago, and that's how I originally got into Bitcoin: I followed those people, and it was one of those things that sounded kind of silly at first, but as you learned more about it, it seemed a lot more interesting.

That's always an interesting feature, right? There are some things you learn more about and they aren't that interesting. Some things are kind of silly—the moon, conspiracy theories that the Earth is flat. I spent an hour one day looking at that stuff, and it's just crazy. I don't know, the moon-landing conspiracies or whatever.

Whereas you dig into this stuff, and you don't have to agree with everything, but there are smart people and it's very interesting. For me, it was 3D printing. This led to my investment in Oculus and Coinbase, which really were both from that thesis—VR, seeing the developers and the Kickstarter community's enthusiasm around what Palmer Luckey was first creating. I also got into nootropics, and that led to investments in things like Soylent. When I joined the firm in 2013, drones were a thing, and we did a few investments around that.

Looking at these interesting hobbyist and hobby communities, there are a bunch of reasons why I think it's an interesting way to look at things.

One is that those are the people who create these things. If you have 20,000 highly interesting technologists, they often build things, right? They're going to build interesting products. It's also a great marketing engine: they're out there, and they often have outsized influence on the internet. They have followings, and they help get the energy going, build things, and market them.

It's not foolproof, and it's hard because a lot of these things end up being niche or don't have—going back to the exponential forces—

You take nootropics: that's still a thing that's around, but I don't think it has created a big tech company, as far as I know.

But I think it's partly because it just has linear forces, not exponential forces, behind it, right? There's not some engine exponentially driving it to have better and better products.

Maybe, actually, though, if you look at a company like Function Health, Function Health is the catalyst for this huge consumer movement around health and the quantified self. Nootropics were a bit of a predecessor to that. In a sense, there is this slow exponential and then very rapid uptake. I think timing is a really interesting question here, because with these movements, you don't know if they're going to play out over 100 years or 100 days sometimes.

Yeah, and you're right. It could just be that 3D printing is a good example. It's still around, but it didn't get as big as people had hoped. I had an investment in MakerBot back then, which was a leader and got acquired, and it's still a hobbyist thing. It's interesting. I think the limiting thing is that, in the physical world, there isn't a kind of Moore's law driving it. That said, I expect that over 50 years or something, it will become a more important thing. You're right—it could just be a timing thing.

Yeah. The vibe-coding thing, to come back to that, feels like this irreversible consumer phenomenon, where everybody is maybe not quite programming but creating software in a way that they weren't 10 years ago. How do you think of that as a decentralizing force? You've talked about the economics of software versus the means of production. The means of production are getting decentralized through these new tools, like Replit and others, and Cursor. Is that sufficient to lead to a renaissance in the open web, or what do you think are the second-order implications of everybody programming?

It's a great question. The thing with the internet and the consolidation—I wrote a book about blockchains, and this was a core theme in the beginning of the book—was talking about what happened with the internet getting consolidated. If you look at metrics like the amount of money, revenue generated, and traffic, more and more, 95%+ of both of those metrics are now in the hands of 5 to 10 companies.

You can make an argument either way with AI. We're already seeing this in the data: a lot of AI obviates the need to click through and go to a website. I think we just saw a report that a bunch of travel sites and others were seeing alarming drops in SEO, which I think is inevitable.

Look, it's a mixed thing. On the one hand, as a user of ChatGPT, it's amazing to just get an answer and not have to go searching again, then go through all these websites. It's this vicious cycle where the websites lose traffic, get more desperate, and put up pop-up ads and other things, so it becomes an even worse experience. This has been going on for about 10 years—a kind of negative flywheel.

On the one hand, it's great for consumers. You get an answer right away. We were investors in Stack Overflow, which got acquired, but I think its traffic has dropped a lot because of vibe coding. Some of the training data probably came from Stack Overflow, GitHub, and other places, but then the technology becomes better. I use Cursor and have used it to do some fun projects. It's an unbelievable tool, and I think it's clearly good for the world, even though it's bad for those websites. It's a great question.

I hope what we're seeing is a renaissance of paid software—businesses that don't need to dominate the internet and be Facebook, but can get to hundreds of millions in revenue. I think we're seeing this, right? From an entrepreneur's perspective, it's a very exciting time. We can see a lot of great products, and I think it's a great time for consumers.

Maybe that will change over time. Maybe companies will need to layer in ads, and the incentives will shift toward things that are more adversarial toward consumers. Right now, I like the AI products because they feel very aligned with users. They're genuinely trying to create great products and charge for them.

Exactly. We sort of call it the emergence of narrow startups, where they charge high prices and deliver exceptional value. Maybe a controversial statement right now is that there are no marketing problems, only product problems, because the technology allows you to be so ambitious on behalf of your customer. The costs actually, ironically, lead to better business models because consumer founders need to think about monetizing early; otherwise, they're just going to go out of business. It does feel like there's a renaissance in paid software that's happening, which makes it a more fun time to build than it was 5 years ago.

Do you think that over time that will potentially shift because people will realize that maybe the low-hanging fruit among higher-paying consumers is picked, and to get the rest you need to layer in different business models—ad-based business models and so forth?

I don't know. No, I mean, it feels like there are so many more consumer needs that are addressable, and they're addressable in such a significant way through the technology that you can specialize and go very, very deep. There is AI therapy generally, then there's AI therapy for people who have ADHD, and then there are people with ADHD who are in a certain life stage and perhaps want to interact in a certain way. You can just go extraordinarily deep. I don't know if it leads to consolidation over time, or if you can continue to specialize and, for a small number of people, be their primary provider.

That might lead to a good topic around the idea maze. You know, Chris, you've talked a bunch about platform shifts. You've invested around platform shifts, and you've predicted them. One of the interesting things about this platform shift is that the properties of the platform are emergent. They're not explicitly defined by Apple, as iOS was. There are things that founders—and even the people training the models—are discovering. Does that change your mental model around platform shifts, and how similar or dissimilar is that to Web3?

Yeah, the idea maze concept originally came from our friend Balaji Srinivasan. I wrote about it a while ago. The way I think about the idea maze is that there was this old debate: with startups, are the ideas more important, or is it the execution? I think the idea maze says they're both important, in the sense that it matters which maze you enter.

I’m entering the AI maze for healthcare, or I’m entering the AI maze for image generation or whatever. Clearly, the idea matters: you go in with an initial product idea. But it also matters that it’s a maze, meaning it’s dynamic. The world will shift, so you can’t predict it.

The canonical example in my mind is Netflix. Netflix started off mailing CDs. The hypothesis was that the internet had changed the way people consume movies, that people would subscribe to them, but that today we needed to send them by mail. Over time, they pivoted to digital distribution. Then they started getting pushback from the content providers, and they pivoted to original content.

They really did 2 almost complete company pivots, but their core maze was right. Their core maze was that the internet would lead to subscription movies in some broad sense, and that was correct. But they were extremely agile with respect to the implementation of that.

To me, that’s the idea maze concept. You’re entering a maze as an investor, and as a founder you need to think, “Am I a person who wants to be in this maze for 10 years? Am I willing to be agile and often persevere through difficult periods?” It’s often emotionally challenging, I think, and not just intellectually challenging. That’s kind of the life of a startup.

Now, when you think about AI, look, we have a very clear megatrend of AI being intelligence. It’s a very broad and important technology. Obviously, everyone knows that. Secondly, you have these scaling laws, which seem to be quite powerful: the models are getting much better.

An important distinction would be that there are specific scaling things, like LLM pretraining, where people may debate at what point you have diminishing returns. Maybe we’re hitting that. I don’t know.

I defer to the experts.

But then there’s that sort of a process, and then there’s the meta-process. The meta-process is AI overall, right? There are people working on reinforcement learning and, I’m sure, 100 different techniques.

AI—the sort of meta-process—means that it’s at this point really an economic phenomenon. There are all these smart people involved. There are business models behind it, and there’s funding. There’s not just 1 process; there are many processes being explored.

It kind of reminded me of Moore’s law. From the outside, Moore’s law, naively—I’m not a semiconductor person—seems like, “Wow, these semiconductors magically get better every 2 years.” If you read a few books about it, from their perspective, they run some fabrication technique, it hits a wall, they freak out, and then some brilliant person from another lab comes up with a new fabrication technique.

Each process would run, have diminishing returns, and asymptote at some point, but the meta-process—the bigger industry flywheel—did not. It led to this smooth growth. My sense is that AI is in that kind of semiconductor-like place, where you have this meta-process that’s very likely to continue scaling exponentially for a very long time.

That creates a huge opportunity for entrepreneurs. The opportunity is obvious: you can build things with capabilities that will grow. There’ll be all these new opportunities and so forth. The challenge is, are the incumbent models going to be sort of God models that subsume your use cases? How do you play that?

What you’re seeing is that people say, “I’m going to go so deep on a domain that that will be my edge. I know everything about this specific domain, and I know that no matter what the incumbent models do, I’ll always be able to have an edge in my product.” Or, “I’ll have such good brand recognition, or a strong user base, or reference selling,” or whatever it might be.

I think that’s both the threat and the opportunity. If you go back—with the semiconductor analogy I mentioned—the canonical case study in Clayton Christensen’s The Innovator’s Dilemma is the hard-drive makers in the PC industry. It was a very fruit-fly, Darwinian struggle where you had thousands of companies and very short life cycles for a lot of them, but also a lot of very successful companies.

It may be a very brutal process for entrepreneurs, in the sense of a lot of competition, a lot of other smart people, and a very dynamic idea maze—but also a massive opportunity.

How do you think, Chris, about native versus skeuomorphic technologies in that context? Everything is changing, especially when you’re building for consumers. Does a consumer change their preferences when these magical new technologies are invented, or, in a sense, does the emergence of native technologies also depend on consumer preferences changing and being informed by these external forces about things like AI?

Yeah, great question. Maybe I’ll define the term first. Skeuomorphic is a term Steve Jobs used with respect to design to talk about how he liked some designs. The original bookshelf app on the iPhone had grainy stuff in the background design, or the trash can on the desktop computer. It hearkens back to a different form factor.

It’s a common pattern in technology and media. When you have a new platform or media form develop, people start off imitating the prior media form. Early films were shot sort of like plays, with a camera and a better distribution model. Then people invented a native grammar of film: close-ups, establishing shots, and all those kinds of things.

A lot of the 1990s internet looked like you would take a catalog—a commerce catalog—and put it online, or take a brochure and put it online. It took 10 to 15 years before you had things like YouTube, modern social networking, and things that really just couldn’t have existed prior to the internet, like user-generated content, where anyone can upload a video.

Some of it is the technology. YouTube couldn’t have existed until you had really wide broadband penetration, so some of it is that the underlying technology takes a while to get there. YouTube also, when it started off, was just funny viral videos. A lot of it was copyright violations. It took a while to develop native YouTubers—content creators.

That’s often just a generational thing. I think it literally is a new generation sometimes: people who don’t look at the technology as a threat, but as an opportunity. That was a big part of it.

Part of it is that entrepreneurs just have to figure out the idea maze. They have to figure out what people want. Around YouTube’s time, there were a lot of debates about whether people wanted to take the NFL and stream it to the web. There were a lot of companies doing that. The assumption was that tastes weren’t going to change. Why would people want to watch 4 people joke around?

Maybe there were analogues, like, “Is that like talk radio, or is that this?” But they really just didn’t understand. I don’t think human nature changed. Obviously, there’s a new generation with different ideas, but I don’t think humans fundamentally changed in a deeper sense. It was about understanding the capabilities of the technology, the cultural shifts around it, and the network effect around it.

With AI, I personally think a really interesting question is—and I’m sure you’ve thought much more deeply about it than I have—whether we’re most likely in a skeuomorphic phase right now.

Right?

What is the native phase going to look like? Usually, for me, at least personally, I like the native phase better because it’s crazier and more interesting.

If you look at image generation, they’re basically taking what illustrators do. But one thing I would mention is a cool thing with photography. When photography first came along, it seemed like a threat to representational painting. You saw art move to more abstract art to get away from that.

If you go back and read things from the time, there was a lot of hand-wringing around whether this was going to cheapen the art form. But an interesting thing happened: a new art form emerged, which is film. You weren’t just copying. In some sense, photographs were the skeuomorphic, quote-unquote, app of cameras, but film was a native one. You had a new art form.

I wonder about that with AI. Right now, you have image generation, which is taking what a human might do and automating it, along with movie generation and the other kinds of videos we see online. But is there a new medium, for example, that hasn’t emerged yet? Maybe it’s virtual worlds or something.

It's probably a bunch of hypotheses as to what it could be, but my experience has been that it's often surprising and hard to predict. That's where a lot of the cool, creative, interesting stuff comes in. It may take another generation, or 5 to 10 years, for a new set of AI-native kids to grow up.

That's right. Yeah. It's actually really interesting because, in a sense, we're in the command-line era of AI. There are some things that you can articulate well with words, but if I describe to you what kind of music I like, it's hard to say—we don't have the language for it. Most people would say, "I like a certain sound with a certain sort of aesthetic, and it's moody, but not too moody, and it's 110 beats per minute." Most people lack the language to articulate the art that they love.

So even the idea of prompt-to-media feels skeuomorphic, and there's got to be a more native way to explore it. I don't know what that looks like yet, but I'd be surprised if it's prompt-based in the long term.

I mean, prompt—I guess people are now calling it context engineering, not prompt engineering, which I think is a nice rephrasing. Some people are, right? I think it's a nice rephrasing because that is kind of what you're doing, right? You're taking all of this stuff I do in the real world that ChatGPT isn't able to see, right? And I'm trying to summarize all that knowledge that's hidden to it—the context—and put it in there. All right.

And that does feel like something that should be automated, right? You have these intelligent machines. I assume that's what people are doing with these potentially new ambient devices people are creating.

Yeah. I mean, even in the media case, my Spotify library is probably much more useful for generating music that I like than my articulation of it.

That's right. You had mentioned in another pod that, if you had 1 issue to get passionate about in the world of AI, it was open source and open-source AI. Do you want to speak to that for a moment?

Well, we were talking earlier about the democratization of the web, or how consolidated the internet or technology is. I think I would argue, and I think a lot of people would argue, that open-source software has been an incredibly important force for democratizing technology.

The reason that you can get an Android phone for $10 and get on the internet so cheaply is basically that all the software is free. Imagine if there wasn't an open-source operating system, and operating system providers used to charge $100. You'd be paying that on the client and maybe on the back end, and there's a whole other stack of software that you'd be paying for. Instead, you're not. Most internet users are using open source; the vast majority of the bits being used are open source.

It also is what makes startups exist, right? We can fund startups, and they can spend hundreds of thousands of dollars, or even less sometimes, and be up and running with really competitive, great software. That's because of open source.

So we think about it a lot, and on the policy side, as a firm, we've been big advocates for making sure open source is around and competitive. First, that means not banning it. There are bills out there, particularly at the state level, that want to put in—not maybe explicit bans, but de facto bans. For example, California had a bill that would have created unlimited downstream liability for software developers, which would have effectively killed open source.

So that's step number 1. Then I think step number 2 is: are the incentives there to create open source? I watched an interview—I think it was a Dwarkesh interview with Satya from Microsoft—recently. It was a really good interview. He argued that open source will always exist because enterprise customers always demand at least 1 kind of open-source alternative. They'll just end up funding it.

Okay.

And that's why you always see this proprietary/open-source combo. But then you have Meta doing this with Llama. I don't know if they'll continue to do that. There are some startups doing it. China has been very into open source. Maybe that's a kind of national strategy; maybe that changes at some point. Maybe you do it at first to create attention and marketing, and then you change it.

The thing with AI that's different from operating systems is that, with operating systems and databases, you just needed a bunch of coders sitting around. With AI, you need massive capital expenditure to train the models. So I just don't know. I think it's an unknown question long term: are there good steady-state funding models for open source? I think a possible outcome, which I think is a pretty good outcome, is that open source is just always a little bit behind, like the way OpenAI is now releasing older models.

Yeah. Yeah.

And I think that's probably a fine outcome. For startups to exist, and for consumers to get inexpensive health care advice, the next-best model in 5 years will probably be good enough for most startups. It will probably be good enough. For the super-high-end stuff, people pay for it. Maybe that's a good outcome—a good kind of equilibrium state. Maybe that's where we're headed. I hope so.

I think it would just be a bad outcome if you had 4 companies that had vastly better closed-source technology and could effectively charge rent to consumers and startups.

Yeah. Yeah. I agree. It's interesting. I think a lot about the early ethos of Android, which felt like it matched Google's open-web mindset. Then, when it became clear that iOS was beating their pants off by being a closed ecosystem, Android became very closed and started to mimic the sort of closed iOS strategies.

We'll see what happens with Meta and Llama, if they replicate that. That's a worrying dynamic. I think the more optimistic case is that we haven't yet seen the same sort of app-platform feedback loop and lock-in that you get from foundation models. There is a case for them to continue to release the next-best model and for the models to be somewhat substitutable for each other.

Yeah, the Android case is a good cautionary tale, right? I think maybe, in some technical sense, some of the code is open source, but de facto it isn't, right? All the services and everything else—you need permission. It was one where they made lots of overtures that way.

That would be the worry, but it does seem—I think it feels a lot better than it did 3 years ago or something, with the China open-source stuff. Yeah.

Yeah, the policy stuff is better. We're seeing the fact that OpenAI is doing older models. It seems like we're in a better spot for open source. I think some of the scaremongering—that a chatbot is going to murder everyone or something—is, like, literally zero people have died from ChatGPT so far, as far as I know. I think maybe people are chilling out on that.

So it feels like we're in a much better spot. I'm cautiously optimistic on open source.

Yeah. 2 years ago, the conversation was a lot about: if OpenAI is the only game in town, over time they take all the economics of the complements. It doesn't feel like that's happened, which is, to your point in that amazing book about blockchain, why there's a lot of interest in acquiring IDEs. They understand that, if the foundation models start to become more interchangeable, they're going to have to move upstream and own user-facing economics.

Amazing. Well, Chris, thank you so much. It's great to hear you talk about consumer AI and all the implications. We're super thankful to have you at the firm.

Well, thank you. Thank you. This was fun.