There's only been one sin, and that one sin is zero-sum thinking. We always worry about, like, oh, is this defensible? Oh, will this layer get margin? Will this layer get value? And the answer has kind of been unilaterally yes. The answer has been every layer has gotten value. Every layer has winners. These markets are so large and they're growing so fast. We're actually seeing brand effects take place. In this phase of model scaling, a lot of the approaches to scaling don't generalize. This gives a ton of room for the application developers to build their own models. I think that right now open source is most dangerous because China is better at it than we are. And as a result of that, ready to go.
Martin, man, I love our conversations. I was so excited when you said you'd join me again. Thank you so much for doing this, man.
So excited to be here. It's great to see you, dude.
I freaking hate these like, how did you get into venture? Intro questions. So, I just want to dive right in. It is a freaking nuts time. So, starting off, how do you evaluate where we're at today in the AI investing landscape? Peak hype cycle, great, super excited, both. How do you evaluate it?
So, I'm kind of two minds. Of one mind is I do feel like my intuition doesn't really work like it has the last 20 years. Um, it's just the future is very uncertain, and one of the reasons is because, yeah, this is really the first time like software development and software creation is being disrupted. And so on one hand I'm like I don't really know what to think. On the other hand, observationally, there's only been one sin, and that one sin is zero-sum thinking. We always worry about like, oh, is this defensible? Oh, will this layer get margin? Will this layer get value? And the answer has kind of been unilaterally yes. The answer has been every layer has gotten value. Every layer has winners. Uh, things that we thought were silly are making money. It's been solved. There's profitable companies. I mean, the business case is there, etc. And so I think the one sin is not playing the game.
Do you agree with the playing the game on the field sentiment? When we look back at '21, you know, I remember everyone saying playing the game on the field. I wish I hadn't played the game on the field. To be transparent, Martin, do you agree that you have to play the game on the field in bunch?
I think behavior should follow business, it shouldn't follow marks. And I think in 2021 behavior was following marks, right? It was like the public markets just decided these companies were valued a whole bunch. You know, Tiger came in with a ton of money and deployed it a whole bunch. And so like I think behavior following investment and marks is a bad idea. But in this case, you have some of the fastest growing companies we've ever seen by users, by revenue. I mean, the amount of value that's kind of shifted to this is so significant. And so, I think investors' behavior should follow that. If not, I mean, what are we doing?
When you think about shifting value, again, I'm diving right in, but this is not going roundabout like a lot of people have fun. And you said about kind of disruption of software development. There is a ton of players in the vibe coding space. They are predominantly all sitting on top of Anthropic. Claude Code is gaining more and more dominance. How do you think about these providers' reliance on a tool that could eventually shut them off?
There are two futures to code. In one future, you've got Anthropic as a monopoly, and another future you have, let's call it an oligopoly or maybe even a bit more of a market of these coding models. And they're just very different futures. And I think when you answer this question, you have to consider both of these. I will say the timing of this conversation you and I are having right now is like pretty soon after Claude 4 launched. And that's like a major model launch. And these models are so episodic. Every time one launches, everybody's like, 'It's the future. Everything's going to happen.' Like remember like the whole Ghibli OpenAI launch and we're like, 'Oh, image is going to change forever.' And then it comes, we're excited, and then it kind of, you know, passes. Um, and maybe that'll happen here, maybe that won't. I don't know. But like for sure like our perception is colored by that launch. So let's consider both of these. I'm going to consider the first one. So historically, models don't really keep much of an advantage because they're so easy to distill. And so we've even in the last week have seen launches of models, um, you know, Qwen and I forgot Kimi, uh, that came out and they're great and people like them and they adopt them. And in that world where you continue to have new models from different providers, you know, I would never count out Google. Their coding models are fantastic. The rumor is that GPT-5 coding is going to be great. So in this world where you've got lots of models coming out from lots of providers, you need to have a consumption layer that's independent, right? And so then all of these companies are going to add that consumption layer value, like for example to non-technical users or to Python users or to professional coders or whatever it is, and that's going to be a very healthy layer.
The other future, let's assume that Anthropic is just a monopoly on coding models. And in that case, you have what you normally have in these situations is they will decide kind of where it's not profitable for them to enter or will change their business model. Like maybe they're like, listen, we want to have the consumption layer, but we're never going to be like an app dev tool company. It's just a different sales motion, a different sales team. And nobody knows where that stops, but they will put pressure on anybody that they view in their core focus. And they will do whatever they can to either capture that margin or to capture that market share. I just think it's just the wrong time to have this conversation right after a major model launch because like I said these models are so episodic and we always assume every time a model launches it's going to be a monopoly and it just really hasn't been the case.
Going to your zero-sum thinking, if you were to put a bet on which future is more likely, which future do you think is more likely?
Oligopoly. This is how the cloud... Well, this is how the cloud played out. I think probably the best analog we have is the cloud, right? You know, the other companies that are behind models can subsidize these things arbitrarily. I think about Gemini and they don't have to do this in a way, um, you know, where they have the same economics as an independent company. And so if you look at how the cloud... remember the cloud, AWS was like 70 or 80% market share early on. Nobody thought they could ever catch up to them. You know, they were the massive market leaders that created the category. I mean, they had way more dominance than Anthropic has now. And Microsoft and Google, like, you know, that's an important big market, we have to be in it, and they just basically spun their way into it. And then you ended up with an oligopoly on the clouds. I see no reason. I mean, Gemini 2.5 is a great model. It's a great model, and if you actually look at it, you know, on the price performance, I would say in many use cases, it's the one that I actually use as my standard model. It's better than Anthropic for some use cases. If you actually, you know, take into account price performance, and Google can arbitrarily subsidize that too. Never, you know, count out OpenAI. I mean, they started the party. Uh, they haven't had a major model release in a while, certainly around code, so that's going to show up. And so I just feel like it's, you know, the players, the money behind the players, the fact that these models distill like this will end up in an oligopoly. But I mean, I don't know, that's just my guess.
To what extent do you think the large model providers in 10 years' time have already been created or are they yet to be founded?
I think that you end up with models with different flavors and there's going to be a lot of new flavor models that will come out. You know, like, we haven't even, you know, Mira and Ilya are out there creating models. We've got these very legit teams that were some of the pioneers. Um, you know, we're just starting up models for the sciences. And as you get more into kind of RL territory, these models really get a certain flavor. They don't generalize nearly as much. And so like that's going to naturally from a technical perspective fragment the models. And so I would say the core base model for like language, search, and code... I mean, I think even code actually, it's still so early. I mean, it's very, very early in the super cycle. In previous super cycles, remember, it took two or three generations for the winners to emerge. I mean, Google was third-generation search. Uh, Facebook was third-generation social networking. Remember, there was MySpace, there was Friendster, and then MySpace before that. And so I think there's a lot of change, there's a lot of change to come. But I do think that both Anthropic and OpenAI have done a remarkable job, remarkable, uh, with brand independence and market share. And so I suspect they'll continue to be stalwarts in the industry.
Are you in either of them?
We're investors at OpenAI.
Oh, you got... okay. Um, my question to you is, you know, fundamentally, um, there's many, but do you think models are fundamentally good investments for venture firms when you look at employee stock compensation and the dilution that comes from it and then the dilutive nature of the businesses?
Yeah, it's a hard sell. Okay, so if there's one thing I've learned, honestly, for anybody that's listening to this, this will be worth like your time. There is no one way to think of AI and there is no like one way to think about models. And the models themselves are entirely different businesses depending on how you talk about the models. So to even answer that question we have to tease apart what you mean by model. So for example, if you look at the diffusion models, like say like ElevenLabs, Midjourney, Black Forest Labs, Ideogram, these are wonderful businesses that have great economics because the models are smaller, the ecosystem isn't subsidized in the same way, right? Like Google subsidizes language and code and video but not speech, right? And so from an investor, these are clearly great investments, uh, because of, you know, if you just look on the metrics alone. On the other hand, the frontier language space, it's much more complicated because there's so much subsidization. Right, right, you have Meta and Google, a bunch of Chinese players that are entering it. So for a subset of the players, and this is why it's a tricky question, for a subset of the players you're like, yeah, clearly these are the fastest growing companies we've ever seen, there's tons of value. These are very valuable entities, right? You know, Anthropic, OpenAI. Um, but at the same time, even three years in, there've already been a number of companies that, you know, have had to exit early. And so I would say it's kind of a high-stakes game where the winners really win, but like it requires a lot of capital to enter the game. And if you're not in one of the leaders, like that, you know, capital is forfeit.
We do a show every week with Rory O'Driscoll and Jason Lumpkin. Rory very aptly, I think, just said, listen, with the transition to AI, every investor's just accepted a willingness to go massively up the risk curve on investing. Do you agree with that?
Well, I think it's the requirement of the game. It's like these are very capital-intensive companies to build. You know, they have to get the capital from somewhere. They're also the fastest growing companies. And so, you know, for the winners, it's justified. And so, I think it's not that investors are willing to go up. I mean, we'd be very happy not to. I mean, I know you would, right? And we'd be great to have great returns with low risk. Um, but the nature of the system and the game which we're playing requires it. And this is what this is, by the way, this is the dissonance in all of this, is just so important to call out, which is on one hand you do have these great businesses that are very fast growing and zero-sum thinking has been tremendously wrong. I mean, Nvidia is continuing to grow in value. The hosting providers, which everybody wrote off as being kind of non-defensible business, continued to grow in value. The model companies, which I can't tell you how many investors wrote off the models. I mean, this question's been around for three years, they continue to grow in value. So every layer of the stack continues to grow in value. So on one hand you're like, it's all working. You should be in the leaders on every, you know, in every layer of the stack. On the other hand, we've seen tons of wipeouts already for the non-leaders. And so it's almost this bipolar or paradoxical situation where you kind of have to play, but it's very, very high risk. And if you don't play, I mean, you're kind of missing one of the fastest growths in value that we've seen in what, 20 years.
Do you think you see the concentration of value to one or two players across markets in every market, whether you look at voice, it's, you know, obviously ElevenLabs, whether you look at it's kind of Replit and Lovable and OpenAI and Anthropic, Cursor?
This is such a great question. So here's one thesis. I, so early we don't know, and maybe in a month all this gets proven wrong, but we actually talk about this a lot internally. And here's one thesis, and this is the one that I'm attached to, which is these markets are so large and they're growing so fast, we're actually seeing brand effects take place. And we haven't seen that since the internet. And by brand effects, I mean if you become the household name, you will get the adoption because it just does not require a lot of education. It does not require a lot of competitive discussion or competitive positioning in the field. You know, I would say for many of these models, I mean, you know, is one better than the other? Yeah, maybe. But they're pretty close. But like people know ChatGPT. It's like it's a household name. My mom knows ChatGPT. Um, you know, people...
When you like, honestly, why did I do Lovable? For the exact same reason that GPT wins. I thought it was the consumer brand that would win.
100%. And I just think these markets are so large, brand effects work. I mean, let's talk about Midjourney. Midjourney was the first that got above the quality bar. It's taken zero investment from institutions. It's still the market leader, um, and it continues to do great. And this is meanwhile while a bunch of other people have entered the market. And so I do think it's not unreasonable to assume that these markets are very large, leaders are going to have brand monopolies and brand moats, uh, and they'll be able to maintain them until things slow down. And in general, I found markets do this, which is when markets are expanding... so markets tend to expand and then contract, right? Think about cloud, right? It was kind of like this funny thing and that became very massive and then of course it slows down. When it slows down then you have the consolidation, um, and then, you know, competitive dynamics come in. I mean, we're clearly in a massive market expansion phase. It's just very clearly the case, and in which case the leaders are going to continue to have, you know, a distribution advantage just through brand recognition.
When does that tail off or does it not tail off? When does the importance of brand and brand recognition dwindle and product prioritization or product quality trump it?
I mean, I think it's as soon as the market growth slows down. You know, let's, I mean, again, let's take cloud as an example where...
Do you think these are actually tools of market... sorry, I'm so sorry to interrupt you, but growth or actually just consumer intrigue? Which is, there's a lot of people who want to try building a website on Replit or Lovable or Bolt or any of them. There's a lot of people who want to try voice with ElevenLabs. To what extent is it market intrigue versus expansion of market?
Well, I just think the expansion of market provides the dynamic so that you don't saturate the user with competing messages, right? I mean, the idea of market expansion is the frontier continues to expand and the first thing the frontier here hears is the household names, and so the household names win. And so I just think that that's a, you know, that's a natural artifact of expansion. And as soon as like the expansion slows, then that frontier is going to hear both names. And then all of a sudden now you're in a discussion of which one to use and not to use. And I again, I think like for the longest time when the cloud market was expanding, everybody knew AWS. It was the leader. It was 70 to 80% market share. And then as soon as that growth slowed down, then all of a sudden market share started to shift dramatically. And it just wasn't obvious. Do you do GCP? Do you do Azure? Etc. But I would say that's less an artifact of the fact that Google, Microsoft decided to enter the game and much more that the market growth itself started to slow down.
So we see market growth slow down and then we see the dispersion of value across players more. So...
That's right. So the market slows down, uh, and once that happens the frontier becomes more saturated, right? Just because we're not adding people as much, and so they will get more of the educated message and they'll start making more decisions and you can have more of a conversation. Like, of course Anthropic would love to have the same brand as ChatGPT as a household name, but how do you reach that frontier, uh, you know, if it's growing that fast? It's just operationally tough to do. Kind of the only way you do it is just through brand recognition, which is kind of this word-of-mouthy type thing. It's like on every podcast and, you know, the friends and whatever. And so I do think we're seeing brand effects happen now, and we saw these in the early internet. Uh, the brand leader tends to get 80% of the market. It just tends to break out for a while and then over time it'll slow down and these things even out based more on product differentiation.
How do you factor that into your thinking when investing today?
You just try to invest in the leader, and it's worth paying up for the leader, honestly. I mean, it's, you know, so I think for me I ask two questions. Question number one is like, for the area that it's focused on, is it the leader? If it is, it's definitely worth paying up. And then the second one is the story actually has been that in a competitive space almost everybody just found kind of a new niche white space. So let's just take the example of OpenAI. I mean, OpenAI was the first to code, right, with GitHub Copilot. I mean, they provided the weights as far as I know, and they lost that. And they were the first to image with DALL-E, and they lost that. And they were the first to video with Sora, and as far as I can tell they lost that. And yet they're still the massively dominant player in language and continue to be so and will be so. And arguably that was the right thing for them because that's by far the largest market by far. And so OpenAI acted totally rationally and has the largest market, but that gave the ability for Midjourney to take image or BFL to take image. Um, you know, Google seems to have grabbed video with Veo 3. Uh, code, I mean, on the model side, Anthropic has, you know, turned that into, you know, this wonderful business. And so when markets expand, not only do you have these brand effects that we were talking about, they also tend to fracture a bunch, and what seems to have been a submarket will emerge as a leading market. And you even see this kind of on the image side, right? You've got a bunch of viable image players that focus on different things, right? Like Ideogram is great for designers and the professional design community. BFL is the open-source community that, you know, especially for developers that use these things in products. And then Midjourney is for, you know, more of the fantasy, like, you know, also professional designers, but it's a very stylized kind of opinionated view. And all of these are independent, you know, viable companies. So I think we're going to see fragmentation for quite a while before we see consolidation.
I need the show to be successful because I'm very open with my troubles. Um, I need your advice. You know, Abridge in the US, I'm not sure if you're in it, but I'm sure you know it. Um, very simple. There's a European player that does like medical transcription for nurses. They went from 1 to 8 million in a year, and we're looking at leading their A. And I'm thinking exactly the same. You're going up against Abridge because you're going to need to compete in the US as this is going to be a big business. Is that a losing game where you are a European competitor?
This is a great question. So, another very interesting thing that we haven't seen in a very long time is we do have geographic biases showing up with AI, and the regulatory environments are quite balkanized. You know, there's language and cultural biases that are also balkanized. And so we're actually seeing a lot of regional players show up. And so I think it's very legitimate. Now, the thesis cannot be European company X wins the American market, but I promise when it comes to AI, the European market is large enough. I promise that. And so I think a very legit thesis is, you know, this becomes a regional player in Europe and then maybe a portion of the US market.
Can I ask you... a lot of people denigrate these businesses that we've discussed because of their margins. They're simply pass-through funnels to the large language models. Do you think that is something that changes over time? And it's the same for all great businesses. Uber started off with terrible margins. Now they have better margins. Same, same.
I just don't buy that these are endemic to the business model. Like, this is certainly not my experience at all. And so there's always this question: if you're a founder and you get access to, you know, relatively cheap private capital and you could do a trade-off between margins and distributions, and it's land grab time, what would you do? And the argument is the incremental user is someone you can monetize forever down the road, and then if you don't get that user during land grab, you could never monetize it. The rational business decision is to sacrifice margin for distribution. It's just the rational business decision, and we've seen this forever. I mean, hell, the web wasn't even monetized, right? Literally. I mean, like this time we can actually monetize these things, but forget break-even or negative margins. It was literally like massively negative because we didn't even have a business model until advertisements came up. So this is like the most rational thing that markets have been doing, at least tech markets, forever. Uh, and it's no different this time with AI. I do think there's a question of, okay, so if you do want to then turn on margins, how do you do it, right? And then you can of course, you'll either have to build a traditional moat, two-sided marketplace, a brand moat, the long-tail kind of integration and domain understanding. So for example, let's say your healthcare company, if they really crack the European market and they understand all the regulation, like Anthropic is not going to take the time to do that, or, you know, so there's clearly pricing power you have on that side. Um, or you have to do actual technical differentiation. One thing that we're learning is in this phase of model scaling, a lot of the approaches to scaling don't generalize. So if I want to be much better at like coding, I may not be so good at something else. This gives a ton of room for the application developers to build their own models that service certain areas that the large models just aren't focused on. And so I think there's even a ton of technical level to differentiate. So my sense is, and this is, I mean, again, I don't want to talk too much about, you know, my portfolio and what I see just because there's sensitivity around the number, but in my experience most of these companies that are like, let's say, break-even margins, it's like a board-level specific choice to prioritize distribution, not just because this is systemically something they have to do.
We mentioned their sovereignty. I am intrigued how you think about safety and safety around AI and models. You've had Vinod Khosla be like, 'We have to lock this down. If this was not locked down, it would be like nuclear secrets being handed out.' I remember then Marc came and was like, 'Fuck that. No way.' How do you feel about the future of safety within this landscape?
I mean, it's crazy to have VCs talking against open source, right? I mean, Founders Fund did too. And for me it's just wild when, you know, pro-innovation, you know, pro-innovation sectors of the economy, academia too, uh, have decided that like open, transparent innovation is somehow an antithesis of safety. And I know that's not what you asked, but like I just want to make the point, it's just, we were in very bizarro land for a while, and it seems like we're coming out of that now.
So let me just draw a bit of a parallel. Do you think we're coming out of that?
I think we're moving more and more into that.
Great. Meta, Alexandr are going to town, Llama fully closed.
Great. So let's go back to that in just one second. I want to answer the question that you, because you actually asked, uh, like a great question on how I view this, and let's go to whether we're coming out or not. So, so how do I think about safety? Um, so I, you know, I was actually very, very close to security during the rise of the internet. You know, I worked for the intelligence community. Um, I worked for Livermore, um, uh, national labs. And then, you know, when I did my PhD, like, you know, a good, you know, 50% of my work was in security. I taught like a cybersecurity policy course. And the thing with the internet is you had these very specific examples of new types of attacks that like impacted nation states, like critical infrastructure would go down. You know, you'd have things like the Morris worm. Like, you know, I mean, you had these really significant examples, and that kind of kicked off this large discussion on how you handle it. And it was so significant at the time that at the nation-state level, you know, we started to think that we have to actually change our doctrine. You know, we were kind of this Cold War era mutually assured destruction. We had to change it to this notion of like defense asymmetry, which meant the more we relied on these things, the more vulnerable we were, right? As opposed to like a country that didn't rely on them because you could be attacked. And then of course kind of the whole terrorist information warfare stuff. And so the implications were so absolute and you had so many proof points and you could articulate them incredibly well. And so if you look at the AI stuff, I mean, for every computer system, you have security considerations, but we've got this 30, 40-year very robust discourse around this that we can draw from and use from. And the thing that I don't understand is how all of a sudden we've decided that these are not computer systems. They don't obey the same laws, and we have to kind of throw out everything that we've learned and kind of like revisit the discourse even though we don't even have the same proof points. I mean, like nobody can make a strong argument on asymmetry or need a shifted doctrine. And if they can, let's go ahead and have that discussion. I still have yet to see the dramatic new attack. It's going to come for sure, but we haven't seen it yet. And so I just feel like the discourse around this is not in line with the reality. It's not in line with historical precedents. And so we should absolutely take these things seriously, but we should draw on the information that we've learned from in the past and the approaches we've taken in the past. The last thing I'll say is the biggest difference this time is in the past the people creating the technology were kind of pro-tech, and the people that were like selling security solutions were like the fearmongers, right? So you'd have somebody create like the internet and they're like, 'This is safe and it's great for everybody,' but then you'd have somebody that create a firewall and like, 'Oh, the internet's dangerous, every sociopath, your next-door neighbor.' So you had both the same voices but in two different bodies based on interests. The interesting thing this time is they're in the same body. So the person that's creating the thing is also like, 'Oh, this thing is very dangerous.' I don't recall the last time we had something like that. Uh, but it's created a dynamic that's just been very confusing for everyone.
Do you not think open source increases the opportunity set for hostile actors like China and Russia to harm us?
I mean, I think it's tautologically true. Like, I think tautologically you can say, do you believe computers and the availability of computers increase their ability to harm us? And I'd say absolutely, computers and availability of computers do. I would say...
Specifically open source over closed source.
So I think that right now open source is most dangerous because China is better at it than we are. And as a result of that, we're seeing a proliferation of Chinese open-source models everywhere. Now, unfortunately, we don't have control over Chinese regulation. And so I would say the answer is yes, because of China, not because of us. And the right way for us to respond is to fuel our open-source efforts, um, against that. So let me just be very specific. So, uh, like I think Chinese open source can be a national security issue for sure, and any of the software that produced by a nation state that we view, you know, quasi-adversarially. The way that we combat that is we also are incredibly open, um, and we also do a proliferation of technology.
What do you think we can learn from China regulatory-wise that would enable us to have the same or better open-source ecosystem/environments?
I mean, to me this is, you know, the United States has a long history of being pro-innovation, pro-innovation for national security, pro-innovation for national defense. I think we should be funding this stuff like crazy. I think we should get the national labs involved. We should get academia involved. We should make this a national priority just like China does. And we should just, you know, a full-throated endorsement of all of this stuff. I think we should do closed stuff. I think we should do open stuff. And we've done this forever. You know, my first job out of college, this is, you know, 1999, was working at Lawrence Livermore National Labs and the ASCI program. And what were we doing then? We were, I mean, the broad program was simulating nuclear weapons. I mean, this is what it was. And a lot of the concerns we have today were concerns we had then around compute. I mean, we actually stopped Saddam Hussein from like importing PlayStations because we were worried about, you know, using them for simulation. We'd put export controls on the hardware and we'd say the same things like, oh, you know, computers out there, like computers, you know, they're going to, um, enable, you know, the enemies and all sorts of stuff. And this is like nuclear weapons. This isn't like some abstract AI thing. This is like actual, actual on-the-ground weapons. The posture that we took at the time and the conclusion is we're just going to be the leaders in all of this stuff, and we funded academia and we funded the labs, and we won, and we were able to control like the technical discourse of the planet going forward. And this time instead we want to put our head in the sand and let somebody else do it. So like they're going to learn from our, you know, our success, and somehow, you know, we're not.
Do Trump's cuts to universities, research labs not impact your ability to do what you just said? Are you not actively going against what you should be doing?
I am very pro-investing in academia and in the national labs. I think there's always a political shift in money depending on what they view is in line with administration politics. Like, I still... Yeah. I can't tell you, you know, I, you know, I did my PhD at Stanford. I've done a bunch of NSF grants. I don't remember ever somebody saying we like indirect costs. Every researcher, every professor, every single one was like, 'Indirect costs are terrible.' Obama, Obama tried to get rid of indirect costs. He's like, 'You know what? Universities, they have a tax-exempt status, so why don't we just have them, you know, spend 5% of their endowments like any other tax-exempt organization?' Um, and you know, that will cover a lot of indirect costs. And he couldn't get it through. So this is a bipartisan issue that is longstanding. And I mean, I would say that like a change is needed now to the extent that, you know, um, you know, I think these things are very hard to implement, but I would say concretely, yes, we should invest in these things. Yes, we need a shift in how funding happens. I do think that like indirect costs have gotten way out of hand, and until it was like Trump doing it, everybody that I know in academia totally agreed. Um, but yes, of course, change and shifts in funding will be disruptive, and so I think all things are true. I just, what I just don't want to do, I don't want to reduce this to a simple like 'Trump does bad things' because I don't think that is the case, and then, you know, 'funding science is arbitrarily good' because I don't think that's the case. I mean, I definitely think we should fund as much or more. I definitely think that shift in funding and a change to the system is needed, and that, you know, the right path through that is complex. I don't quite know it.
You very kindly said that I asked a good question on the reversion back to closed source when we mentioned Alexandr joining Meta, what it meant for Llama. I said quite zero-sum-wise to your point, we're clearly seeing a movement back towards closed and away from open. How do you see that and do you disagree with my statement there on the transition?
No, I think that's so... I agree on the ground 100% that I think we're seeing a movement away from open source, but the rhetoric around open source has shifted, right? I mean, we just had the AI, uh, what was the name of the bill that just came out? I mean, it's like the American AI policy and recommendations is a full-throated endorsement for open source. So I think discourse-wise, there's more support for open source than ever before. I think ecosystem-wise, I think you're right. I do think it's quite likely that we're going to see less open source. Now, listen, OpenAI has said that they're going to open source. That would be wonderful. And if they do that, I think that would be very, very positive. Um...
I just, I have no idea. I hope so. It'd be a very rational... I mean, here's the great, maybe here's the, like we say open source, but it's such a misnomer when it comes to AI. I mean, the standard model of open-sourcing AI is you open source the smaller model and you keep the more capable model closed source, and it's a way that you get distribution and brand recognition, but you don't actually erode your business. Uh, this has been very, very successful as a business model. And unlike actual software open source, just because you release your model doesn't mean somebody can replicate it. Like to replicate it, you'd have to like recreate the data pipeline and the training pipeline. And so, you know, I think that there's just like a lot of concern of investing, you know, hundreds of millions of dollars or billions of dollars to train something and then just giving all of that away. But I feel very confident that the business justification is there, and behavior will always follow business, and we're going to continue to see open source be a large part of the ecosystem. And remember, historically, open source has only been about 20% of the total market value. I would say it's much higher than that for AI. So in a way, we're doing better than software has historically.
What did you believe about the AI landscape that you now no longer believe? We've touched on so many different elements. My mindsets have changed around so many. I mean, the one for me that I've just consistently got wrong is just how fast these coding models advance. And this is probably just sunk cost fallacy. My entire life, I've just been this nerdy programmer. I've been programming since the '90s. I mean, it's like it's my happy place. And I just never thought that they would advance to the level that they have. I mean, I still develop most evenings, and it's just, you know, instead of watching a sitcom, I just goof off and mostly writing like old video games or whatever just for fun. Like, it's silly stuff, and I'm already at a point that I just couldn't... I just couldn't go back to working without them. And I've spent, you know, 30 years without them. And it's just their ability to offload all of the stuff I didn't want to learn is remarkable. Like the thing that kept me away from code for a while, which is I would kind of dabble with it, I would drop it, is you have to just learn all of this, like all these weird frameworks, and like it's none of the knowledge is foundational. It's just like some random dev came up with some weird way to do something, and you've got to kind of learn, you know, some poor design decision to do it, and none of it made any sense. And it just felt like you're wasting your brain space on poor decisions made by random open-source developers. And that was programming in the past. Like, program in the programming. So, let me...
Just put it in context. In the late '90s, programming was: you download your IDE, you'd sit down at your computer, you'd program something, and then it would turn into a binary, and then you'd run that binary. And so, like, you could really get a lot done just by sitting down and writing code. You know, I would say like 2015 or so, writing with something, it's like you'd have to download like 50 million packages, and to run it you got to run some stupid dev server, and to actually have anybody else use it, you got to learn how to host it. And you know, it was a bunch of libraries that were dealing with incompatibilities for all this weird platform. So like 90% of your time had nothing to do with code. Like 90% of your time is just dealing with all the environment platform stuff.
And so what's so nice now is you can just focus on your code. So like now I literally just — I mean, I use Cursor and I just have the AI tell me how to host the thing and tell me what package to use and whatever, and I just strictly focus on what I want and the logic. And so it's almost like it's brought coding back. And you can see this across the industry. Like, I've got — I mean, I grew up in the industry. I know a bunch of very strong developers that have been developing for a very long time that have basically stopped — they're like running companies now or whatever — and they're all back to programming at night. And I really think that, you know, how there's like the adage of like, I don't know, the old man that goes into the garage and makes the train set for nostalgic reasons. I think the modern version of it is these old systems programmers, you know, vibe coding at night just because it's become pleasant again. And so I know you asked about the thing that's kind of surprised me the most, but I really think it's such a marvel what these coding models are able to do, and they add very real value.
Do you think they make 1x engineers 10x or 10x engineers 100x?
10x engineers 100x would be what I — but I don't actually think it's that. I think they make 10x engineers 2x. I would say every company I work with uses Cursor, right? And then if I actually look at has that increased the velocity of the products coming out, I don't think that much, just because so much...
So what's changing then? Because dev productivity is going up, so is the quality of product going up if the product release cadence isn't?
I just think the things that are hard remain really hard. And so, you know, let's just talk about creating a model. So let's say I'm creating a new model, a new frontier model, right? And to create that new frontier model, I've got to collect data and I've got to run a pipeline and I've got to sit with my Jupyter notebook and I've got to look at the loss curves and I've got to rerun it. And that's just a lot of experimentation and so forth. You know, there's no coding model that's going to do that for you. But if I wanted to create tests or a test suite or visualization or write documentation, it's actually really good at that. And so I would say that probably in the long run, having more robust, maintainable code bases with less bugs is just as likely to be the impact as feature velocity. Because you know, in startups — again, I'm an infra guy. This is probably different for the apps. Like I've always thought apps had no technology to begin with. Like every time I look at vertical SaaS, I'm like, why do we even care about the technical team? It's CRUD, man. It's like CRUD is create, read, update, delete. They all do the same thing. They all just kind of look like a web app. They're all like, who cares about the technology? The technology is simple. These are all these kind of go-to-market things and whatever. But infrastructure is different. And infrastructure is like very real trade-offs in the design space that only somebody that understands computer science would know. So for infrastructure companies, I think it's quite unlikely that AI will really help speed that up because it comes down to something that the developer has to decide on, has to articulate the trade-offs. But I do think it could really help with the development process so you have less bugs and things like that. And so I actually view it more as a more robust developing methodology that doesn't necessarily speed up the core product.
Given the kind of dev productivity changes that occur because of these tools, how does that impact defensibility within companies today? If time to copy — which Misha at Five said this on the show — he said time to copy has basically been reduced to nothing. To what extent does that change defensibility for companies?
I mean, I still think we should just go back to the split between apps and infrastructure. For apps, how long does it take to copy it anyways? I mean, you know that there are entire companies whose stated purpose is just to copy another company in the app space. It's just so easy to do. I mean, there is no core technology for a random app. There's no differentiable technology for a random app. Let's say that you're creating, I don't know, some healthcare vertical SaaS thing — like you could contract and you have been forever. The actual app — I mean, the business is actually the long tail of understanding that domain. So I just don't think it changes that paradigm at all. And then when it comes to core infrastructure, which is what I focus on, things like databases, foundation models, there's no way that right now models can just copy. And the reason there's no way is it's not that the models aren't capable of doing the technology. It's just that there is a long tail of understanding of the trade-offs for the particular use case and domain. And because it's a new market often, then you understand that through market exploration. And so I just don't feel that — I think these models really help with the software development process for non deeply technical areas like apps. Sure, they can help speed it up, but over time, all of these reduce to a long tail understanding of the market. I mean, Aaron Levie said it so beautifully. I mean, do you know what the average — what do you think the average PR, pull request, is for a production codebase? Like how many lines of code is the average change that gets accepted, would you guess, for like some production enterprise app?
It's two. It's two. Yeah, it's very, very small. It's actually two, but let's say it's 12, right? And what does that two or 12 lines signify? That two or 12 lines signify probably some learning in the field or some understanding of what is needed. And so the long tail — the thing that's the hard thing is to understand the specific deployment environment and market you're going to. That's the hard thing. The hard thing isn't the two lines of code. That's actually quite easy. And so in many ways I would say, you know, the AI is getting rid of the middle, right? Like so very new computer science, like models they don't know how to do just because nobody's done it before, and that's kind of pushing the state of the art. And then in the app space, all of the hard stuff is the business anyways, right? And this is why the changes are very small and you learn everything through go-to-market, which the models don't know just because you're exploring a new market. And it's all the stuff in the middle that they're helping us with. And so, you know, for me it's just kind of net creative.
Do you think that CS holds the same weight as a study in education discipline that it always did and you would always recommend it, or does that change in a world that's bluntly more democratized in terms of creation like we discussed?
I mean, I feel very strongly that if you care about building systems out of computers, you have to understand how they work.
What do you think we do today, Martin, that we will look back on in five or 10 years' time and go, I can't believe we did that? It could be prompting. It could be choosing the model that we're working on. I find it ridiculous that we are supposed to choose which model. Like Grok 3, Grok 4, Grok 5, Grok shopping, Grok weather. What the — just figure it out.
Well, I'm just taking it from a programmer's view. I mean, I just think hopefully we'll just stop worrying about frameworks altogether and maybe even languages, maybe even a proto-language evolves and we can just focus on logic and fundamental trade-offs. I mean, we've gotten in this very backwards world where these days programmers think about all the non-fundamental stuff and they don't think about the fundamental stuff. Let me give you an example. So I always worried — this is going to be this weird philosophical rant — but I always worried, you know, while I was doing grad school and when I was doing research, that we kind of entered a space where there's so much research that has been done over the years that you never know if you're doing something new. Like you just couldn't do the literature search. There's so much. And so the entire industry just spent all of its time redoing research, you know? It's like you're cleaning a room and you're trying to sweep out the dust, but rather than sweep it out the door, you're just kind of moving it. Like you move it to the bed or you move it to the wall, and then that's all you do is you just kind of sweep the dust around, but you never actually get it out of the house. That's what research felt to me. It was like we're in this mad delusion. And on top of that, it also felt like many of the most important problems were kind of between disciplines. And so in order to even solve them, you just have to know too many things, and we couldn't do that. And so I just felt like the entire scientific industrial establishment was just kind of redoing the same stuff. And so in a way I think AI has the ability to pull us out of this mass craziness, this mass ineffectiveness, which — a, it's very good at telling you if you've done it before, right? It's very good at that. It actually knows all the literature, knows all the history, and it's also very good at tying different disciplines, right? And it is an expert in all of these things. And so I think we've been stuck in this morass, and it's a bit of a liberator so we could actually focus on the new problems and know we're doing new things. And so I've got this very optimistic view of where it's pulling us. And so I know it's more of a philosophical answer to the question that you asked, but in a way I think it needed to happen to get to the next level of problems that we need to solve.
In terms of societal implications there — I mean, the worst question ever is like, oh, the job displacement question. But I am intrigued because on the one hand I see intense job displacement happening faster than ever, and then I'm also very aware of Brad Feld wrote a brilliant post where he basically said every single cycle, every time, we've always said, oh, what are we going to do — calculators, what are we going to do — computers, what are we going to do — AI now, what are we going to do. To what extent does this actually require the 'what are we going to do' versus another? For God's sake, don't we see the pattern?
