OpenAI vs Anthropic vs Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning

Matan, it is so good to have you in the studio. You’ve just insulted my continent with the suggestion that we’ve only come up with bottle caps while you came up with Transformers. Not wildly untrue, but this is going to be a fun show. Thank you so much for joining me.

1. Will AI actually increase GDP?

Thank you for having me, Harry. It’s a pleasure to be here.

I was just doing a show yesterday with Rory and Jason, and Rory was basically saying that the fundamental question is: will we see an increase in GDP coming from AI and the coding developments that we’re seeing? Will it lead to GDP increasing above the 2% average for the last 200 years? Do you think we will see meaningful productivity gains from the AI tooling that we’re seeing, or are Uber’s concerns validated?

I think, yes, absolutely, we will see tremendous growth from these tools. I think it takes time to permeate through because you can tell, on an individual basis, almost on a problem-by-problem basis, that we can solve problems faster with these tools.

Companies generally organize around solving problems. If you’re organized around solving problems and you have some set of personnel, you might say, “This is the number of problems we can solve at a given time based on how many people we have.” Everyone is now going to be able to solve more problems with the same number of people, or solve the same number of problems with fewer people.

2. Smaller teams or bigger ambitions?

But it takes time for the resource allocation to adjust. A lot of businesses will have to ask, “Do we want to solve more problems now because of the increased leverage that we get, or do we want to solve the same problems but do it in a more efficient manner?” That is a question that a lot of businesses will be grappling with.

Do you think we will have fundamentally smaller teams, which ultimately suggests that option number 2 is the one that most people take? Or do you think we will have the same-size teams and just go after a more expansive area?

It’s really not obvious because there are dynamics that are hard for me to predict. But what I will say is, again, bringing it back to problems: all of these companies are now going to have to think, “Okay, we have all this new leverage. Do we want to solve the same problem? Do we want to increase our ambition and solve a bigger problem? Or do we want to solve more problems that our users maybe have?”

I was watching Andrej Karpathy, and he was talking recently about how the 10x engineer is actually wildly misunderstood. You won’t see the 10x engineer; you’ll actually see a smaller number of 100x engineers and the rest. Do you think that is the right way to look at the future of engineering talent?

3. Why agency matters more than credentials

Directionally, yes. What is a 10x or 100x engineer? I don’t necessarily agree with the language around it, but I think it implies 10x of what? Pure output? If, when you say 10x, you mean how much code they’re writing, I can now write a billion lines of code with these tools. It might be terrible lines of code, though.

Maybe the way that I like to think about it is as load-bearing individuals in an organization. It’s kind of like, if you remove this person, things fall apart. In some organizations, there might be people where, if you remove them, nothing happens, and they’re not load-bearing in that case.

Basically, these people who have very high leverage are now being handed a tool that gives them even more leverage. Using the language of 10x or 100x, yes, they’re levered up and they can have even more impact. I think, with that leverage language, those who know how to use leverage will be able to have even more impact, and those who don’t will, on a comparative basis, be that much less valuable to a business.

When we think about the 2 different options you mentioned—that you can do more with the same-size teams, or reduce teams and do what you already did—if I’m thinking as a leader today, what would be your biggest advice to me on how I should think about resource allocation for tokens internally?

4. The resource allocation problem: tokens, dollars, people

Yes, this is a great point. This resource-allocation problem of tokens—it’s not just tokens; it’s dollars and tokens and people—is, I think, going to be the thing that over the next 24 months every C-suite is going to be thinking about.

I think the right way to go about it is: what is the core competency for our business? What actually matters for the business that we are doing? Then, how do we allocate resources accordingly?

In other words, if you’re a logistics company, your core competency is probably not software development. You might have added a lot of software engineers as a means to an end to deliver on your logistics goals, let’s say. But that might not be your core competency.

What you should be thinking about is not, “How do we get more engineers to make more features?” That’s what engineers have in the past been judged by: how many features did they ship in a quarter? Instead, it’s, “What are the actual output metrics that matter for our business, and how do we now allocate resources—whether it’s dollars, tokens, or headcount—to more dramatically move the needle on that business outcome?”

I think this is great for the world because part of the reason why so many organizations got so bloated is that we were in a period of time where everyone was focusing on intermediate metrics. If you’re an engineering team, we wanted to ship 3 features this quarter. Did you ship 3 features? “We shipped 4.” What a great quarter. That doesn’t necessarily matter for the business at all.

Now it’s finally coming back to what matters in the first place. What are the business metrics that we want to move the needle on? Is it customer satisfaction? Is it revenue? Is it market share? You can tie every individual’s work back to that, whether it’s marketing, sales, engineering—all of it.

5. Kirkland's $500M AI bet and the build vs buy question

Kirkland announced a $500 million spend. You’re friends with Winston from Harvey, a fantastic guy, who obviously—I’m sure you have spoken about this, although I don’t know if you guys have—has a big spend: $500 million across 5 years to internally build their own Harvey or Legora. How did you think about that?

I mean, it’s interesting. Talking about core competencies, Kirkland is spending half a billion dollars to build its own AI tools. My understanding is that building AI technology is not a core competency of that firm. I was surprised to see it.

Now, I actually think this is good for Harvey because there’s nothing like trying to do something yourself to make you realize, “Oh, this is actually really difficult. This doesn’t actually matter for us—to have the in-house ability to build this ourselves. Let’s go and have someone who is an expert in this build it for us.” That is my sense.

My favorite is also the number of people who say, “See, we told you how easy it was.” You’re like, “It’s so easy.” They’re committing half a billion dollars. That would suggest the opposite.

I had Brendan on the show the other day, and he was fundamentally saying that the next 12 months would be the most value-accruing 12 months for AI infrastructure companies. We would see the models, the products, and the AI application-layer companies be most at risk and denigrated. Would you agree with that?

I would disagree. I’d pretty strongly disagree for a couple of reasons.

One, actually sticking with the Kirkland example, I think that we’re so used to a world where the moat in software was, “I know how to do this and you don’t, so you’re going to pay me because I have the engineers who know how to build this, and you simply cannot.”

Now, the world going forward is going to be one where there is nothing that no one can build. Every single piece of software anyone will, in theory, be able to build.

Now, going back to the resource allocation, is it worth your time and energy to go and build it, or should you go to someone else who has already built it or can do it faster?

To me, an example of this is: suppose we had a very busy day at work. I could probably go and pick up lunch for everyone on the team. I know how to do it. I know how to walk out the door, place an order, hold the bags, and bring them in. Just because I know how to do it, is that an efficient use of my time? Probably not.

I’m probably going to say, “You know what? For my resource allocation, I’m going to pay someone to go and do that for us,” because at Factory, our core competency is not that the CEO goes and gets lunch for everyone. I think it’s somewhat similar here: just because you can build a lot of these things does not mean you should.

And in fact, oftentimes you want to be really ruthless about the few things that you and your team own and do end to end. If it’s not relevant to your core business and your core competencies, outsource it.

What would you like to do, but it’s not a core competency for you, and so you don’t do it because of focus?

I enjoy making breakfast. I haven’t done it in 3 years. There’s nothing wrong with it; it’s just not time-efficient. It doesn’t make sense to spend my time doing that.

All right. Okay.

But it is something that I do enjoy. To your point on models, applications, and infrastructure, I’m not sure if you’ve seen the meme of the Microsoft org chart. It shows different segments, and they all have guns pointed at each other, just to show that in Microsoft there’s a lot of bureaucracy and everyone is fighting over who gets to do what.

I think that image is pretty accurate to what’s happening right now with the models, the application companies, and the infrastructure companies, where everyone is trying to commoditize the other. Everyone is trying to say, “Oh, no, this one is irrelevant. All the value is going to be here. All the value is going to be there.”

The reality is that value accrual is a time-dependent phenomenon. It’s not like there’s one player whose steady state gets all of the value. That’s not how it works. Maybe for this next year, this person has the pricing power and gets the value; in the next period of time, these people get it.

We are all, whether overtly or not—and maybe I’m saying the quiet part out loud—trying to commoditize the people that are not us. For example, we’re model-agnostic. We want to give our customers the best pricing, the best performance, and the best speed for whatever task they want to do in their software development.

We want to make sure that OpenAI, Anthropic, Google, and Microsoft are all under pressure to provide the best models as cheaply and as quickly as they can, and don’t feel like they can just charge whatever they want. Similarly, the model companies want to make it so that the applications are trivially easy to build, and that the product is really the model.

6. The bear case against Factory

The infrastructure companies have their own spin on this, but the reality is that everyone is trying to commoditize the one that’s not them. From Microsoft’s perspective, it’s very much in their interest that models with access to proprietary data get differentiated value and capture a huge amount of value, because that validates their business model. It’s a big push and pull over who can get the leverage.

What is the belief that would invalidate yours?

The bear case against Factory is if one model provider gets significantly better than all of the others. A key thing for us is that all the models are going to be roughly as good as each other. One is a little bit better at review, one is a little bit better at testing, one is better at Python, and so on. It all fluctuates every week.

Even already, people have a hard time keeping track of which model is number 1 and what the latest thing is that came out. If one model ends up going way above all the others, that’s a case where companies might want to go completely in with that provider. But then that’s a monopoly for the entire economy to be worried about.

Is the rate of model development sustainable? What I mean by that is, I was with the founder of Nebius the other day, and he was talking about how every few weeks we see new models. I said, “You’re wrong. It’s every few days.” Especially when we look at Chinese open source, it’s like 3 or 4 a week.

Is that rate of model development a feature of the time that we’re in, or is it an ongoing characteristic or trait of this environment?

I think eventually we’ll stop seeing them as model releases, and they’ll feel more continuous, just like before when it was GPT-2, then GPT-3, then GPT-3.5, then GPT-4, GPT-4.1. Then you get more and more decimal releases, like 4.5 and o3. Eventually, they’re just not going to announce it. It will be, “Hey, look, here’s our model that’s continuously getting better.”

People are already experiencing fatigue. Engineers at the enterprises we work with can’t keep up with every single model that comes out, nor should they. That’s the whole case for the application layer, whether it’s us, Harvey, or whoever else: we’re going to figure out what model is best for what use case, where the trade-off is between cost, quality, and speed, and deliver that to you based on the task that you have.

7. The rise of open-source models

It’s hard to focus on what matters for your business and also keep track of all these models that keep coming out.

A big question that people have is around the rise of open source and whether everyone is concerned by the amount they’re spending on tokens being so much larger than they thought. “Hey, we spent our annual budget and it’s May [__]. Maybe we should move to open source.”

We’re seeing more and more great companies use frontier models, see where they can get to, and then move to open source to get as close to that as possible. How do you feel about that being a considerable threat to the market for frontier models?

I think it’s a really important counterbalance, because it allows you to make trade-offs over what tasks you want to put what level of intelligence on. A lot of enterprises will realize that many of the tasks they’re doing don’t need the very frontier models. They can do them much faster and much cheaper with these open models.

Again, it’s part of resource allocation. To do good resource allocation, you want to be able to be anywhere in that cost, quality, and speed trade-off.

I love the GIFs on Twitter where it’s me naming a file and it’s like the massive cigar with the blowtorch. I love that.

No, because it’s such overkill. But there’s also a funny dynamic that emerges, which is that there’s kind of an ego thing: “Oh, no, the work that I’m doing can only be handled by a frontier model. This mere open model can’t deal with the work that I’m dealing with.”

Even when I first started switching over, I’d think, “I don’t think an open model could handle this.” Then it’s like, “No, it probably can.” It’s a funny thing to mentally deal with—deciding manually, or having the router do it for you.

My question is, enterprises like security. They like reliability. They like ease.

Yes.

When you have frontier models that are packaged perfectly, priced clearly, and secure, won’t they just go for that—the easy option—over trying to be smart and intelligently route to different open models?

A couple of things. It’s easy when there’s only 1 of them, but as we said, a new model comes out every week. If you have to go through the full enterprise process to get every new model in, it’s not very easy.

It’s also really expensive. If you’re seeing your costs go up like crazy and you don’t have an ROI case, it doesn’t make as much sense.

Something interesting is that we’re seeing 3 phases happen in these enterprises. Phase 1, a couple of months ago, was the board yelling at the CEO: “Hey, Mr. CEO, what’s your AI strategy?” The CEO says, “Shit, I don’t know.” Then the CTO says, “What’s our AI strategy? Let’s make sure we adopt AI.”

Phase 2 was AI at all costs: token-maxing, making it part of your performance reviews. “We’re going to measure how much you guys use AI. Everyone, you have to adopt it.” That was about getting as many people to adopt it as possible. Phase 2 happened a lot faster than people might have expected.

8. The AI spending hangover

Now we’re entering Phase 3. Phase 2 was the debauchery—the long night, taking shots, having a great time, using all the AI. Phase 3 is the hangover, where you look at the bill and say, “Oh my God, we are spending so much. I have no idea what the ROI is. Is this helping our business?”

That’s where a lot of these companies are now. I think this is why routing is so important, because they’re realizing—and this is a true story—that one of the CIOs I was speaking with had realized they’d been spending hundreds of thousands of dollars per month on people asking Opus 4.8 questions like, “Hey, how’s it going?” “What are my macros from the food I ate today?” “What’s the weather like?”

It’s like, guys, we don’t need the frontier of human intelligence to be doing this stuff for us. In some cases, it’s not even work-related.

Will we see a contraction, then, given that the hangover period is being realized?

We might see a short-term contraction in usage of the very frontier models, but I think it’s healthy. It’s healthier to do that than to be blind to it and then have a sudden, dramatic change.

Uber announced last night, I think it was, or yesterday, that they were having a $1,500 budget per individual. How do you respond to or think about that?

We’ve literally seen this with dozens of our customers. Initially—and this is a lesson for our post-sales team—we came in and said, “Oh, by the way, we have these user limits, and here are the models.”

“Go crazy.” This was before we had routing, and it happened a couple of times with customers where the usage would go crazy. They hadn't spent the time to actually determine which parts of the codebase they wanted to dedicate these tokens to and which they didn't.

Then they were like, “Oh my God, we're spending so much. This is crazy. We need to put in token limits.” The first time this happened, we were like, “Oh my God, their usage went down. What's going on?” But spending time with them, we realized, wait, we need to make sure with every customer that we're having a very clear conversation with them: it looks like you guys are spending a lot of tokens on some of these things. Have you thought about consciously saying, “Yes, we want to do this”?

Sometimes we'll proactively set those user limits just so it's better to be aware as you're going up, as opposed to just going crazy and then realizing. What's happened with Uber publicly has happened privately with a lot of customers of ours. There's a little bit of shock where it's like, “Okay, wait, let's put in these user limits.”

9. Token spend as a % of dev salary

But then you come to a question of, well, wait, this team is really important. They should have a different user limit than that team. We're just getting towards this world where you have very nuanced resource allocation throughout your org.

10. How the coding market matures

To me, the biggest question that I ask myself—and I think we need to ask ourselves as an ecosystem today—is: if Marc Benioff says that he spends $300 million on Anthropic for his devs, that is 3.8% of salaries. Okay, great. What will that number be in 3 years' time? Because if it's still 3.8%, fuck. If it's 20%, fuck—again, but fuck positive.

And if it's Brendan at Mercor who says that he's spending more on tokens than he is on headcount, fuck again, but even more positive.

What do you think that percentage of dev salary is in 3 years?

I think it's actually a more nuanced question than we might think. I actually think it can be as low as 0% for some individuals, and it can be as high as thousands or tens of thousands of percent for some individuals.

And what's the dependence there?

It depends on what the unique skills of those individuals are. I'm saying individuals and not devs in particular because I think the way we even organize roles is going to be very different. I'm not sure “dev” as a word makes sense.

Traditionally, it's like custodians of code, right? The people who do anything relating to code are engineers or developers. I think everyone is going to be loosely interacting with code in your org, whether they're in sales or marketing.

But I think the difference is there are going to be certain people where, again, we're coming back to resource allocation, who get more leverage by using more tokens. Then there are going to be certain people who don't really need tokens at all, and that's not how they deliver value to the business.

For example, maybe our best salesperson is best used not by having them use tokens, but by having them go and meet people face-to-face. That's an obvious example because they don't write code in the first place.

But similarly, maybe there are some engineers who actually do their best work by spending time with users, spending time with their customers, or doing some data analysis that's not very token-expensive. Then there are going to be others who are delegating to dozens of Droids in parallel, working on a ton of different crazy features, refactors, and migrations.

I don't think it's going to be a consistent number across the board. In fact, I would argue that if your org has a standard number where it's like, “We want every engineer to be at this percentage of their salary in token use,” you're probably painting with way too wide a brush.

If I were to say, give me an average number, what will that average be? What will the median be?

I would say order of magnitude. It'll probably be comparable to salary.

Comparable to salary.

Yeah, on the same order of magnitude.

Within a 3-year timeline.

Yeah.

What percentage of tasks today using frontier models could be done with open-source models?

Probably 80% to 90%. It's typically the planning that really needs the frontier models.

But if it's 80% to 90%, does that not just present the biggest bear case ever against Claude Code? You're just taking away 80% to 90% of that time.

Well, it depends, because that 10% to 20% could be the most important tokens. It's maybe 10% to 20% of the tokens, but those are really, really important because they're kind of decision-making tokens, perhaps.

Sure, but it's very similar to how we structure human orgs, which is that oftentimes leadership makes very key decisions that determine the fate of the company, and they don't spend the most hours. If you look at the human hours of a company, most human hours are not spent on making the decisions. They're on gathering data or implementing things, but then there's a select few hours where it's like, here is where we're going to make this irreversible decision on the strategy. And those people who make those decisions are also typically paid a lot.

Yeah, but the assumption there would be that you'd have to increase that spend for that 10% even higher.

And that's what's happening already. The frontier models are sometimes getting more expensive, or you're using the ultra-high reasoning, or this type of thing.

If this planning thing is the very key thing, we'll spend on it, but it doesn't necessarily mean that most of your tokens are going there. It's just for certain key steps. Maybe you want to spend a lot, and it's worth allocating the budget there.

But then, once it comes to, “Okay, we have the plan. Now let's implement it,” the open models are typically really good.

11. Factory's controversial culture: sales and engineering as one team

When we think about what you're willing to spend on, I asked Brandon how much it costs to hire great AI researchers, and he was like, “Tens of millions of dollars.”

Yeah.

Have you found the same, and is it impossible to hire great AI researchers in competition with Anthropic and OpenAI?

We are a very opinionated organization, and the people who like the opinionated stances that we take are willing not necessarily to go and try to maximize the dollars that they can get out of the market. That said, it is still pretty competitive.

What do you think is the strongest opinion that you have that most people disagree with?

I would say the opinion that we have that I think, in the space that we're in, is the most controversial is the way that we treat what product is at Factory.

I think there are some very commonly held beliefs at the labs or at some of our competitors who are also doing software development. Growing up in the Bay Area, there's a very common Silicon Valley fallacy, which is that research is the pinnacle. Then there are engineers who implement the research. They're not quite there, but they're still great. Then there's sales and marketing and all that dirty stuff.

“Oh, if only we could build a better product, it would sell itself, and we wouldn't need to deal with sales and marketing.” It's just completely delusional.

The product at Factory is the entire journey from the very first time they hear our name till their 10th renewal after a decade of being a happy customer. The software is a big part of that journey, but so too is the marketing that we do and the people that we have running that. The same goes for the sales process: the people who present themselves in discovery calls, demos, or solution engineering are all part of the product.

That entire thing is the product. Everyone is first-class. It's not like we have engineers who are wholly at the office and you're not allowed to speak to them. No, no, no—we have engineers and salespeople sitting next to each other. There are no engineer corners, sales corners, or any of that stuff.

When salespeople close a deal, engineers say, “We closed a deal.” When engineers ship a feature, salespeople say, “We shipped a feature.” It is entirely one team, entirely cohesive. Everyone is first-class. There's no first class or second class.

This is shockingly controversial, in particular in the Bay Area and in particular in coding or in AI. It's messed up. It's so messed up.

I think the reality is it will come to haunt some of these companies one day, because right now, where there's a gold rush and everyone's desperate to sign and get more tokens from these people, it's easy. In my mind, it's kind of like they're astronauts in space where there's no gravity. Your muscles will atrophy. Gravity will come back.

If you don't have a good sales and marketing team because you don't give it respect, the second gravity returns, all of your muscles will be atrophied and you won't be able to compete. I would say this: name a legendary company that has a shitty sales or marketing team. You can't.

No, but I can name companies that have shitty products but great sales and marketing teams. That's the ironic thing on the flip side. In fact, it seems like many more legendary companies have shitty products but great sales and marketing teams. I'm not going to name them because I'm going to jumble them. I'm sure we're thinking of some of the same ones.

Yes, if Chad Peters were here, he would say them.

Yeah, that's right.

Does what it takes to be a great engineer change when you essentially become a prompter and manager of agents versus a creator and doer of tasks?

Yes, it very seriously changes. This is actually why we're selecting for this very intentionally. The culture that we just mentioned is really important, because the best engineers are going to be the ones that don't see sales and marketing as dirty work but as, again, an important part of the product.

As an engineer, your job is no longer just to ship a feature. You're owning full end-to-end outcomes: here's the way the customer is behaving; here's how maybe we can change that behavior to make them a better user long term. It makes them more agent-native. They get more out of our product, and we can then follow them through that journey.

You enable the salespeople so that they know how to talk about it or how to demo it. This is a full-stack engineer that goes way beyond just engineering and into sales, marketing, enablement, and all that. Those are the parts of engineering that really, really matter. Those are the parts of engineering that have typically made engineers good founders.

The parts of engineering that become less important are, funny enough, the things that Silicon Valley has really bragged about a lot: competition-winning or Olympiad-type achievements. Are you as fast as possible at coding, or do you memorize all the different nuances of these different languages? Those are the parts that don't matter. If you memorized some coding language or some syntax of a coding language that someone else didn't, it doesn't matter.

People like the credentialism, but I think they misunderstand the VC mindset right now. As a VC, I'm happy to share how we feel. Fundamentally, there is intense uncertainty around what Anthropic and OpenAI will do and who they will kill at the application layer. In a world where we desperately seek certainty, we look for validators, and the validators of someone being a Math Olympiad winner, or whatever that is, serve as a good crutch in the wake of not having other certainty.

Yes, but it's a crutch. It's helpful. It's a good indicator. Generally, you can't win competitions if you're dumb, right? It's pretty rare.

However, for these types of engineers that we're looking for, that's cool, but it's kind of irrelevant. What have you built? How have you taken ownership and agency of things end to end?

This is not to say that we don't have people on our team who have won Olympiads. I think they're great, it's fantastic, and a lot of my friends have done it. But there are also certain high schools that really focus on, "You must do the Math Olympiad. You must do the AIME to then go to the IMO." That's the path to success, where actually that's kind of an anti-signal because you're not owning your fate or choosing your agency. You're going through the funnel.

But then there are people on our team who are from the middle of nowhere, where no one else in their high school ever did this stuff. They took the agency to say, "I think this is really fun. I'm really competitive. I want to compete at this," and they went and did those competitions on their own. Those are the cases where the signal is still positive for this kind of engineer of the future.

I don't think I've told you this, but do you know who you always remind me of?

Who?

Matt Damon in Good Will Hunting.

Oh, that's—I mean, I'll take that to the bank.

Have you not been told that before? Let me say that one more time: you look identical.

That's very—

We're going to put up an image here, and you're going to do a side-by-side. All right, we have a clip. This is the clip. This is the starter to the show. They're going to be like, "Uh-huh. I totally get it. That makes absolute sense."

Can I ask you, when we go back to what makes developers great and how we think about structuring the team, what role does not exist today that you think will be incredibly common in the next few years?

I think it's starting to exist more and more, but I think it's this GM, or general manager, role for someone who used to be an engineer, where you own an outcome end to end that is not just a shipped feature but a business outcome.

Even at Factory, we have this now, where there are people who will own the marketing copy if they're going to be releasing something. They'll own the outcomes in the product metrics, and they'll own enabling the salespeople. It's way beyond what a typical engineer does, and it feels like, again, owning more of a business outcome—more entrepreneurial, higher agency, just spreading their reach.

I think that's true of every function. Believe it or not, I sometimes post on different social media platforms, and I just said my biggest advice to any students here would be: just be full-stack in whatever you do.

If you're doing marketing, create the copy, make sure that it's ready to post, post it at the right time, and amplify it. You have to be in every element from start to finish—not just, "I do the copy, then hand it over to designers to create the visuals, and then they hand it over to a social team."

12. The age of the polymath is back

Is that not just the same for every function? We're expecting everyone to be full-stack in every function. The age of the polymath is back.

Growing up, I was obsessed with math and physics, and I was so jealous that hundreds of years ago people like da Vinci, Euler, or Newton could be polymaths. It was because their fields were relatively shallow: chemistry wasn't that built out, mathematics wasn't that built out, and physics wasn't that built out. In da Vinci's case, art, engineering, and sculpture weren't that built out either. So you could get to the frontier of multiple disciplines within your lifetime.

Then, growing up in the early 2000s and 2010s, pre-AI, fields were so deep. In my case, theoretical physics and string theory were so deep that you could spend literally 50 years catching up on all the literature and academia that had existed before you contributed anything new. This was infuriating to me. It was so frustrating.

With AI, we're now completely the opposite. These tools can get you up to speed with the frontier, obviously with a lot of uncertainty about certain details. You won't have the depth of other people, but they'll get you to the frontier way faster than ever before.

If you're someone who's good at thinking around constraints, thinking about systems, holding uncertainty in your head, and being okay with that—knowing there are unknowns and knowing that you can still push the frontier forward despite that—you can be a polymath.

You can push forward and create innovations in how to do developer marketing, while at the same time pushing forward the frontier of token caching for software development agents, at the same time as being an incredible solutions engineer. These are things that you can now do all at once, and this is something that's very top of mind for me.

In our hiring process, we want to find the people that can be those polymaths. The era is totally back. Polymaths are back.

I've had a lot of people say that agent operations will be the leading function that doesn't exist today and that will be very common in 3 to 5 years. Do you agree with that?

What is the definition of agent operations?

Agent operations is the creation of agents and the maintenance of them. It's being able to go into different functions and say, "Social media: I'm going to create agents that allow you to create, distribute, and share posts. Marketing and design: I'm going to create agents that allow you to create visuals, share them amongst each other, edit them, and collaborate on them."

13. What we'll look back on in disbelief

To some degree, everyone should be able to do that on their own. But I could imagine a world where there's someone whose job it is to find places that aren't as efficient and, similar to operations now in organizations, make them agentified. They're using agents to make the organization more efficient wherever possible.

But I think in general, if you have people in certain functions who aren't proactively doing that, it's probably a bad sign.

What do we do today that we'll look back on and go, "Oh my god, I can't believe we did that"?

For an engineering team, it's writing release notes. That's crazy—that people used to spend hours writing release notes or writing documentation.

Not everyone knows what release notes are. What are they?

It's basically cataloging the changes that you've made in the last month or so and sending it out either internally or externally to your users.

And documentation, generally—Stripe has a really great reputation. They had incredible documentation. So many APIs had horrid documentation; Stripe was the pinnacle. They were so good at this and spent a lot of time doing it.

14. Models, apps and infra: who gets commoditised?

5 years from now, it's going to be, "Oh my god, I cannot imagine, cannot believe that these people who get paid so much money spent hours of their time doing this." I think that's something that we definitely won't do.

Does that reduce the impact of Stripe's great documentation if everyone is equalized?

Yes.

But I think Stripe has plenty of places where it can differentiate, and I think it's a better world where everyone has documentation as good as Stripe's.

Totally agree with that. How does the product-review, and especially the code-review, process change in the next few years?

Yeah, I think what's cool about this agent-native software development is that review has been a big problem. Basically, the first phase of rolling out AI coding tools was, “Oh my God, look how much code we can generate. It's incredible. I'm generating a ton of stuff.” Then phase two was some poor staff engineer who had to review hundreds of these slop PRs that didn't adhere to your standards, were completely misformatted, and all this stuff.

What's great about having this kind of full end-to-end software factory, as it were, is that it's now very clear that there's an ROI in investing in things that make your agents more ready for production. Examples of this are making sure your agents have access to up-to-date documentation, and making sure agents can spin up a remote machine so that they're not just generating the code, but can actually run it, see what the outputs are, and iterate based on that to make sure it's actually good.

Setting up things like CI/CD, good linters, or good pre-commit hooks—these are all things that the best organizations in developer experience would invest a lot of resources in. But they would do it because it makes it easier for engineers to work and easier for them to onboard.

The impact of doing that well is just one-to-one correlated to how many engineers you have. With agents, though, the impact is now 10x or 100x, depending on how many agents you're using, because the better your DevEx, the better your agent ends up adhering to your standards. That means there's less time that that poor staff engineer has to spend reviewing your PR, which means faster throughput in your software development.

When agents are the buyers and you're selling to agents, how does the world change, and does the value of a great API increase?

I think that value is increasing, especially because the things that make it easier for agents tend to be the same as the things that make it easier for humans. At some point, in theory, that could change, where you're actually training models or training agents to be as efficient as possible at communicating with each other. But the downside there is that it's not as human-readable.

If you think about agent-to-agent communication, agents don't give a damn about UI or design, but they do fundamentally care about data structures, potential integrations, and documentation. Do you know what I mean?

Yeah. Yeah. Yeah.

So, I think one thing is that if you don't have careful standards in place, it can get bloated pretty quickly. But I think the best organizations that are the most agent-native actually put in a lot of guidance, like, “Here's the UI side of things and how things need to be,” while being very aggressive about pruning anything that's unnecessary.

15. Why the company is called Factory

Making sure there aren't bloated, gratuitous comments in all of your code—there are ways around it. But that's kind of where the human's job changes a little bit. Their job goes from—part of our name is Factory; part of why it's called Factory is because the future of software development is where these organizations, instead of having engineers who build the software, are going to have engineers who build the factories that build their software.

Visually, whenever I say this, I always think of Tesla's factories. I don't know if you've ever seen videos of the inside of Tesla's factory. It's all these robotic arms moving around, and you have the assembly line going through. There might not be as many humans on that assembly line, but you know damn well that humans designed this process to optimize the throughput and produce more Teslas in this case.

16. Labour displacement and the problems AI will finally solve

In this new world of software development, human engineers are not going to be involved as much in writing the actual code. But they're the ones who are going to be involved in figuring out how to make sure it's not just creating all this bloat, or technically getting the job done and passing tests, but doing it in a way that isn't dramatically increasing technical debt. They're kind of building the scaffolding around this factory that produces their software.

Do you worry about labor displacement when we move from working in the factory to working on the factory?

Short term, yes. Long term, no.

Short term, yes, because it's just a shock to the system. There are all these big layoffs happening that are pretty aggressive, and these are thousands, tens of thousands of people who had a job and no longer do. So, I think that does worry me.

Long term, though, I am very much not worried, because the reality is that there is a huge number of problems in the world—a ridiculous number of problems in the world. A large percentage of them can be solved, or can be helped, with software. Very few of those problems that can be solved with software are we currently solving with software.

If we're going to be flooding the job market with tons of engineers, that means we can now allocate them across the broader economy to solve more of these problems in the world. If we have more engineers who are solving more problems in the world, that is a net good.

What problem is not currently being solved with software that will be enabled by this new technology? Because everyone's like, “Climate change,” and I'm like, “Great. How many people have I found doing climate-change technology well?”

None.

Yeah, well, maybe part of that is because all the Googles have been hiring all these engineers. Distributing great engineering talent to more problems, I think, is going to be a good thing. The economy has to match, though, and properly incentivize them, and that's something that I think will take a little bit of time. That's the intermediate period.

But so many health problems—so much of pharmaceutical research can be advanced with better engineering. The thing that really upsets me with some of the people who are talking about pausing AI development, or any of this “It's a bad thing and it's going to harm society” thinking, is that dementia is a go-to example where everyone understands how big of a deal it is.

That is something that can be solved with better AI and better software. It's a matter of time. We will solve it, and we can solve it. By saying you want to slow down AI, you're saying to people who have relationships with loved ones who have dementia, “No, no, no, sorry. You guys have to maintain that relationship for a little bit longer. We're scared. We don't know about AI.”

I think it's pretty harmful and pretty selfish to say that. To me, it doesn't make sense. It doesn't make sense.

Do you agree with government intervention?

In what capacity?

In free markets, when you think about the allocation of resources, there are times when it is suboptimal, from a human-morality and societal standpoint, to see engineers at Anthropic working on optimizing Claude Code when they could be working on optimizing healthcare systems or more critical, mission-critical things. Governments can intervene immediately, offer subsidies, and offer economic incentives. Do you agree with that, or do you believe in Adam Smith's invisible hand?

I think it's certainly useful in some cases. I don't think anyone would argue that the government should never intervene ever in the economy, because there are some things—especially as they relate to military uses, safety, or weapons—where you're definitely going to need some involvement.

I think some incentivization can be helpful, just because there might be some problems for society that capitalism doesn't see the immediate feedback loop of. You might want to juice the incentives a little bit to get an outcome that you're looking for. But generally, I'm pretty reluctant. I think you need to have a very good case for why you need to do that.

Even with the example of climate change, it's obviously a very sensitive or important subject for a lot of people. You could make the case that the faster we develop AI, the sooner we solve climate change, because AI can help us solve a ton of these problems. But to develop AI faster, you might need to consume fossil fuels and emit them, and that emits CO2 into the atmosphere.

17. Are we in an AI bubble?

The question is whether, in the short term, it might be slightly worse, but it ends up getting us to solve the problem way sooner instead of dragging it out over 50 or 100 years. There are some of these cases where the natural free market will incentivize it the right way, and there are some cases where it won't. But I think you need to be very, very careful about the cases where you do want the government to say, “Hey, we want to step in here.”

Do you think we're in an AI infrastructure bubble?

Maybe there are some short-term blips, but long term, absolutely not. Not even close.

I think there might be similar corrections to the thing at Uber, where we realize, “Oh, we were going a little haywire. We weren't allocating it appropriately,” and it's like, “Okay, let's lower consumption a little bit.” But on the net, absolutely not.

What bottleneck do we have today that will be completely solved within a few years?

I think the biggest bottleneck by far, working with all these organizations, is the human side of it. It's behavior change, especially—

What you're saying there is selling into large enterprises and how they do change management.

Yeah. Or even on an individual level: if you're an engineer who's been an engineer for 30 years, it's hard to change those patterns. You're stuck in your ways to a certain degree.

But there's also a funny thing where some of these engineers who have been engineers for a very long time, or who have been engineering managers, might be more reluctant to use these tools, but sometimes they're better because they know how to delegate. They know how to deal with some of the junior engineers where, if you tell them the wrong thing, they're off in the cave doing the wrong thing for 7 days, and they come back with something completely useless.

On the other end of the spectrum, there are people earlier in their careers who don't have as much of a standardized workflow that they're used to. They're more eager to adopt these new workflows, but they don't know how to manage people. They don't know how to delegate as well. So there's an interesting balance there.

18. Lessons from selling to enterprises

When you look at now, you sell to some of the largest enterprises in the world, in some cases. What do you know now about selling to large enterprise that you wish you could tell young Matan 2 years ago?

This is the first job I've ever had, which I think is always a funny thing to say, because prior to this I was a theoretical physicist. I literally never had a job—never worked in a coffee shop, any of that. I had literally never been paid to do anything aside from physics until this, which is a whole separate thing.

I will say the thing that has been the craziest learning—and this is obvious to anyone who's in sales, or to Chad and Chris; to them, it's obvious—is that the thing that was the most visceral, altering thing was meeting people face-to-face. It makes such a big difference if you're trying to sell them something.

But also, you should never try to sell something. You should always try to understand their problems and see if the solution that you might have can actually help them solve that problem. If you go into a conversation trying to sell something, especially to engineers, don't waste your time. If you go in trying to have genuine curiosity about it, it's really easy, because these organizations do their engineering so differently.

I find it fascinating how all of these different banks, consulting firms, and pharmaceutical companies have such different ways of building software. It's really interesting to go talk to them and understand it, and the best way of talking about it with them is face-to-face. People love talking about their problems, and they love talking about all of the bureaucratic nightmares that they have to deal with.

By understanding all of that, you can get a sense of whether our software is a good fit for them and whether it will help solve their problems. It's also just so fun to meet up with them a year later and be like, "I remember when you had to deal with that shit, and now you don't have to." That's such a rewarding feeling—making their lives better in that way.

19. From string theory to Factory: the origin story

In terms of being there in person and the sales process, you got Sequoia very, very early on. Sequoia is obviously one of the best and most prominent investors. Can you tell me the story of how you got Sequoia, having never had a job and only being paid to do physics?

I was obsessed with physics basically since I was 12 because I was a bad student. My geometry teacher told me that I had to retake geometry in high school. I never tried in school, but I always prided myself on being good at math. When she told me that, I was like, "Are you kidding me? She thinks I need to retake geometry? I'll show her."

My first order on Amazon ever was textbooks for Algebra II, trigonometry, precalculus, Calculus I, II, and III, differential equations, and maybe a linear algebra textbook. I bought those textbooks, and then, the summer between middle school and high school, I studied all of them and did all the problems in them. In high school, I took exams to place out of all of those classes.

Then I asked my dad what the hardest math was. He said string theory, which is technically physics, not math, but I was like, "Okay, I'm going to be a string theorist." That was literally all I cared about for basically the next 12 years of my life. All I cared about was math and physics.

I ended up going to Princeton because they had a great physics professor I wanted to work with. He's a famous professor named Juan Maldacena, and I was the first undergraduate to work with him and write a paper with him. Then I ended up coming to Berkeley to do my PhD and work with a great adviser there.

Only at Berkeley did I realize that it all kind of came crashing down: "Holy shit, I've just been doing this because it's hard and because someone said I couldn't do it. What the hell do I do with the rest of my life? This is crazy." Everything came crashing down.

Of course, that crashing-down moment. Why did it take so damn long?

12 years.

12 years. You're slow.

I have tunnel vision. When I get obsessed with a problem, it's all I think about.

I was in law school for 2 weeks. It was a quick realization.

Some people are faster. I wasn't quite as quick.

Honestly, part of it was that, as a grad student at Berkeley, you have to teach classes. I was teaching a class to 18-year-olds who didn't give a shit about physics, and I was like, "Oh my God, this would literally be the rest of my life—sitting and doing lectures and teaching these classes as a professor." I think I had a 1 out of 5. I was horrible. It wasn't a good fit.

It was this existential crisis: What do I do? I realized it was probably going to be either quant finance, which is what a lot of math and physics people do, big tech, or startups. I ended up doing the quant finance interviews, like every good physicist does. I almost took it and almost went to New York to do it.

At the last second, I had an adviser I spoke to who was like, "You know what? Stay at Berkeley for a bit. Don't do it. You're always going to be good at math. You could always go and do quant finance. Stay at Berkeley. Explore, learn some stuff, whatever." So I was like, "Okay, fine. I'll do that."

20. Discovering code that writes itself

I ended up taking my first computer science classes at Berkeley. I had learned how to code for physics—for simulations and all this stuff—but never in a formal class. I'm very competitive, and I found that in these classes I was doing better than some of the computer science students. That was very competitively satisfying. I was like, "Oh, okay. I'm going to take more of these."

It wasn't until I took a seminar in what was called program synthesis at the time—we now call it code generation—that it completely nerd-sniped me. The idea here is not machine learning for video, audio, or images, but code, with the explicit purpose of creating itself. There's something so fundamental about that.

A decade of physics—physicists and mathematicians are never interested in the case of n = 3 or n = 4, four dimensions. It's always, "What is the n-dimensional solution? What is the arbitrary, fundamental solution to things?" There was something so fundamental about this idea of code generating itself, and it got me obsessed.

I stayed at Berkeley, and for the next year that was what I spent my time on. My adviser was very chill and allowed me to take AI courses. Eventually, I realized that the way to actually solve this problem was not in academia but in industry, and to properly solve it in industry, you'd have to start a company.

But I knew nothing about starting companies. All I cared about was math and physics. I didn't know anything about this. So what does someone who wants to learn about starting companies do? They order Peter Thiel's Zero to One on Amazon and look up on YouTube how to start a company.

I read Zero to One. It's an incredible book. I know it's so clichéd, but to someone who didn't know anything about that—and, growing up in the Bay Area, shockingly, I just didn't care about any of it—reading this was so concise and beautifully written. I loved that.

21. The cold email and 3-hour walk with Sequoia

Then I was watching a lot of Y Combinator videos and all this stuff. I stumbled upon a Stanford VC Club podcast with a guy whose name I recognized because, at Princeton, when I wrote that paper with Juan Maldacena, I had cited one of his papers. He was a theoretical physicist, and I remembered his name, but he was on this podcast talking about how he sold a company for $1 billion and was an investor at a place called Sequoia.

He also seemed pretty sociable and normal in the video, which—I don't know if you've interacted with—

I'm not sure, dude.

Theoretical physicists, though, compared to other theoretical physicists, he could maintain eye contact. He was somewhat normal. Very rare.

That's such a low bar.

He could hold himself in a social setting.

Yeah. So I was like, “Okay, who is this guy? I had to talk to him.” I ended up writing him an email: “Hey, I’m Matan. I also used to be a physicist. I wrote a paper with Juan.” I didn’t say the last name because if you know, you know. I would love to get your advice. He responded that day and invited me down to Sand Hill, and it was supposed to be a 30-minute meeting. But we ended up going on this walk, and it ended up being a 3-hour walk.

On this walk, it turned out we had very similar reasons for getting interested in physics and very similar reasons for leaving physics. At the end of it, he was basically like, “It was great to meet you, Matan. You absolutely need to drop out of your PhD, and you should either join Twitter right now because Elon just took over and it’s hardcore for your résumé if you voluntarily go there, or you should start a company.”

I was like, “Okay, thank you so much. I appreciate you taking the time. I’m going to go think about it.” But in the meantime, I had already known about Factory. It was just that I didn’t want to ruin the meeting with a pitch.

You didn’t want to transactionalize this.

Yeah, because it was so incredible. We had the exact same reasons for getting interested. I didn’t want to dirty it with a pitch.

No, it’s kind of like an LP where, at the end, you’re like, “I don’t want to ask for a check.”

Exactly. The crazy thing is, the next day I went to a hackathon in San Francisco and saw across the room this guy who also went to Princeton, whom I recognized but didn’t know super well. I ended up talking to him. He was also interested in this problem. We joked that it was like intellectual love at first sight. This is my co-founder, Eno.

From that day forward, we spent every day talking nonstop. I had some shitty demo that I built. Eno is 1,000x better of an engineer than I ever will be, so for the next 72 hours, he and I put together this better demo. Then I called up this investor and was like, “Hey, I have something cool I want to show you.” We hopped on a call, I showed him the demo, and I was like, “What do you think?” He was like, “Eh, it’s okay.”

I was like, “Are you fucking kidding me? This is going to change the world. What are you talking about?” He was like, “Okay, well, would you work on it full-time?” I was like, “Yeah, absolutely.” He was like, “Okay, drop out of your PhD and send me a screenshot.”

Keep in mind, my parents immigrated from the Soviet Union to the United States with basically nothing. The fact that I was doing a PhD was their pride and joy. It was the thing they were most proud of.

22. Dropping out and the $1M check

There was so much momentum. He answered my email, we got along well, and I met the co-founder the next day. I was like, “You know what? Fuck it.” I dropped out, sent him a screenshot, and he was like, “All right, you have a meeting with the Sequoia partnership tomorrow morning. Be ready to present.”

You’ve never presented to a venture partnership before. So what happens?

I made a shitty deck.

We go to the Sequoia HQ and put some slides together.

Keep in mind, I didn’t even know who the hell they were. It was just like, “Oh, these random people. Okay, whatever. Yeah, I’ll go talk to them.”

I wish it was recorded because I’m sure I came across as so arrogant.

How did it go?

I thought it went fine. They asked some questions. Again, I didn’t know anything about VC land or startup land or any of that stuff. Retrospectively, I know Alfred, Pat, and Roelof were all in there. They were all asking questions, and I was probably dismissing some of them: “Oh yeah, we’d solve that easily. We do this, we do that.”

Keep in mind, this was in April 2023. This was way before anyone was thinking about agents, way before people were even using Copilot. We were talking about fully autonomous software development agents. It was kind of a blur.

The next day, Shaun called me and was like, “Hey, we want to give you a check.”

How big was the check?

A million dollars. And do you know what he gave me for it? Do you know what the terms were? 5% post.

What? I’m not being rude. Why did they bother doing a partnership meeting? In the nicest way, that’s like a coffee. I know it’s a dick comment, but when you manage 7 different time zones, it was a different time. Early 2023 was a different time.

20% post.

Yeah.

Just for listeners, on the last funding round, that would be like a $300 million position, not including dilution. It was one of those things where a lot of people I spoke to were like, “You should go shop that around. You can get better terms because it’s Sequoia.”

When you have a connection like that, there’s a certain thing to me where, obviously, you want to maximize the position for the business, but no one else would have believed in me except him. No other partner would have understood. I had literally never had a job before.

Retrospectively, it’s like, “Oh yeah, no one else would have done it.” Trust and loyalty and belief matter so much more to me than the price tag you get or whatever. I want to make sure that the people I have in my corner are there because we’re building a legendary company. It’s not just going to be 10 years. This is a lifetime.

Would you tell founders to take a discount for Sequoia?

So generally, yes. They’re the best firm, in particular if there’s a special connection with you and the partner or a special reason why them in particular. But I think what really matters is that you want to have people who are there for you when the days are tough and when it’s not obvious.

When you’re a hot company raising a hot round, everyone’s your best friend. It is their job to make you feel special, and they are really good at it.

Well, it’s the best way someone’s tried to woo you.

There are just some people I don’t even know. I don’t want to name names. There’s this one investor in particular who’s still in the game but more of the old guard. I’ll say that much.

I remember beforehand, people told me, “By the way, he’s really good at making you feel good about yourself.” I was like, “Yeah, whatever. I can deal with that. That’s fine.”

I remember leaving the meeting being like, “I’m the fucking man. This is my destiny. I’m going to build a legendary company. I got this.” Then, 30 minutes later, when it wore off, I was like, “Oh my God, he got me. He did the thing. He made me feel special.”

A lot of investors, when a company is hot, are going to do that, and they’re really good at it. That’s why they’re great investors.

For me, what’s really important as we’ve built out our board in particular is having people who have deep conviction when it’s not obvious. That’s what really matters, because when a company’s hot, everyone’s going to be excited. It matters when it’s not and there are going to be tough times. How do they behave then?

How did you get Ivanka Trump as an ambassador?

Through one of the best hires that I’ve ever made at Factory: this woman, Francesca. The way Francesca and I met was at a random conference. I was seated next to her and Alex Pall, who’s one half of The Chainsmokers.

Obviously, people know them as The Chainsmokers. They’re also incredibly good investors, which sometimes surprises people. Francesca and I got along quite well, and, weirdly enough, we also grew up in the same hometown, which was another kind of weird coincidence.

In the process of them wanting to put a check in, the way she did diligence and the way she carried herself made it so clear that she was a killer. They wanted some allocation, and I was like, “No, no, no, sorry. It’s going to be this.” She was fucking relentless.

She came to our office and was like, “Hey, we need to get to this much. How can we do it? If I do this and this and this, the business value that we provide to you is going to make it worth more than giving that allocation to someone else.” She was kind of hounding us.

We were having a conversation, and I was like, “Look, Francesca, if you want more ownership of Factory, you could just join us.” I said it kind of as a joke: “You could just join us if you want more. This is the highest we can do.” But then we both were like, “Oh, interesting.”

We talked about it a bit more and realized, “Wait, this is an incredibly strong fit.” We ended up bringing Francesca on board. For Alex, it was kind of tough because she was incredible and they were very close. Alex has since been happy because she’s helped us deliver a lot of returns for them.

23. Does Ivanka Trump add value as an investor?

We’re the biggest fans of theirs, and we still have a very deep relationship. She was very close with their firm, Affinity, from her investing days. Then we were introduced, we got along really well, and that was how the connection was made there.

Does Ivanka Trump provide value? People will look at her and be like, “Her branding’s just a name, whatever.”

And I don't mean that disparagingly at all. I think people often think that with famous celebrity names. Does she actually provide value?

Yes, she is. First of all, she's one of the kindest and smartest people that I've met. There are people that you meet who are famous and are kind of a letdown, or you think, “Oh, they're different from what I expected.” She is genuinely so kind and so intelligent, and people throughout tech and throughout the world really love her, and for good reason.

She has an incredible network, and she's so generous with her time. There is dirty-work investor help that she provides that some other investors who are more known as investors do not do. So, she and the firm more broadly really earned that right on the cap table.

That's really good to hear. I hate the statement. I'm not sure he was quite your hero, but people say, “Never meet your heroes. They always disappoint.” I think that's just total bullshit.

Yeah, I remember meeting Doug Leone, who was one of my heroes, and he did not disappoint. I left thinking, “God, he should have been even more of a poster boy for me because he was so great.”

I totally agree with you that that's very funny. I would love your thoughts on some market composition that I'm struggling with. When you look at Cognition, Claude Code, Codex, and Cursor now with Grok, how does this market evolve and mature? Is this an AWS versus GCP? Is this an Uber versus Lyft? What is the mature state of this market?

Yeah. I think what is necessary for the best outcome for consumers is going to be models that are separate from the applications. You, as a consumer, do not want to use applications that are provided by the same people who are giving you the model, because the incentives are misaligned.

The incentives are misaligned. Why?

Because, if we take the example of coding, if I'm a model provider and I'm working with a large enterprise and giving you a coding tool, I want you to use as many tokens as possible because I'm an API business, and I get more money the more tokens you use. I don't have a huge incentive to be more token-efficient, other than wanting to give you a good product experience, but that's not a strong incentive.

Versus if you have model providers and an application layer that allows that enterprise to decide between the different providers, if you're a model provider, you better damn well be the best, the cheapest, or the fastest, or else you'll never get tokens through to you. So, it puts the best incentives on the model providers. There's that independent agent—in our case, Factory—and then that gives the best prices to the enterprise.

It also gives them the best flexibility. If one model is really good at this language or that language, it allows them to adjust between them. In a world where you're vendor-locked in, you can slowly get lazy and see slower shipping, and as a consumer, you end up getting a worse experience.

Okay, so it's not good for the consumer if the model is tied to the application. Okay, cool. But bluntly, we're seeing Codex and Claude Code eat a huge part of the market. What does the market look like in 3 years in terms of market maturation? This is going to be different from cloud.

I think a lot of people suffered because the cloud providers came and said, “Hey, look, sign this 3-year deal. We're going to give you a big discount. We'll get everything good for you. It'll be all right. Come on in.” Then they would do that, jack up the prices, and once you're standardized on one, it's going to take you 2 years to switch to something else. So, good luck. You're stuck with us, and we're going to charge you more.

Everyone has scars from that. Now, every CIO I speak to is really keenly aware that we cannot throw our lot in with just 1 model provider. We're going to need to be agnostic. You could be agnostic by saying, “Hey, every engineer, we're going to give you Claude Code and Codex and Gemini CLI and all these other tools.” But then the problem is that you're asking your engineers to use 10 different tools.

Or you can use someone like Factory, where you can use 1 tool and decide, kind of like in an auction, on a task-by-task basis which model provider you want to use. Do you want to use an open model? Do you want to use Frontier? Which one of those?

How can you help me understand the paradox of, “Hey, we need to be more cost-efficient with Replit, but we're going to run the same prompt on 3 models at the same time”?

And I don't mean that there's no diminishment to Replit. They're providing a great product, but I also haven't seen them, or that use case, as much in the enterprise. I could see it for consumer use cases, where you're not as cost-sensitive because you're not doing things at crazy scale, where it's kind of fun to see, “Oh, I wonder what Gemini does versus OpenAI versus Anthropic.”

For some enterprises, if there are things that are very sensitive or very secure, sure, you might want to do that. But for a lot of use cases, if you're a nontechnical person building an internal dashboard, you probably don't need 10 different models to generate different iterations of it.

Totally get that. In terms of market maturation, what happens to the Lovable and Replit market? We saw OpenAI release a competitive product last night, and I just don't know what happens there. Can you help me understand it? It's not obvious to me.

Part of it is because not too many people close to me use those tools frequently. Most of the people that I know either don't use AI tools, or they're technical and using Factory. Also, none of my friends use Factory. What do we—come on, we wouldn't be friends. So, I need to understand a little bit more about that user.

My sense is that we're still in the early innings, so I'm sure they're quite agile in figuring out the exact niche that they want to occupy. But it's not super obvious to me what the focus is, because my understanding is that some of them have been pivoting toward the enterprise a little bit.

I think, from the enterprise perspective, they've been pivoting toward the enterprise in non-developer-centric functions. If I'm Lovable, I'm going to sell to sales teams, marketing teams, and customer support teams, to allow them to create amazing materials with no experience developing.

24. The coming security danger zone

Sure. In that case, I think that niche does make sense a little bit more. I think it would be ill-advised if they were to try to go to the niche of nontechnical people writing code for code's sake, because I think that is going to be run by—if you're going to need enterprise controls over who has access to what databases and what code and all that stuff, that's going to be run by the engineers. That's going to be where Factory goes.

If it's things like a salesperson wanting to build a customized demo app or customized website for something, I could see, in some cases, that having some value there.

Are we entering a danger zone for security? A huge amount of net-new code is being created that may not be as secure as previous code was, and we're seeing just the worst hacks and security leaks. This is just the start.

Yes. Yeah, it's going to be crazy.

When you say it's going to be crazy, what does that actually mean? The amount of code that's being generated?

The code being generated is growing exponentially. Security efforts aren't growing in kind, and so I think there's a lag there. I think there are probably going to be, in the next couple of years, some pretty big incidents that occur. There probably generally have been some already; I just think whatever incidents have occurred, no one is going to admit it, or typically they'll be reluctant to admit whether AI was involved or not.

25. Should US startups use Chinese models?

But also, I think we haven't even seen the most adversarial behavior yet. People can use these tools to be quite adversarial, and so I think security—the higher the stakes, the more important it's going to become. The security part of the market is really important.

Do you think US startups should be allowed to operate so extensively on Chinese open-source models?

Yes. Using an open model is fine. I think there are 2 concerns. One concern is if you're sending your data externally to a different nation, which is one concern. The concerns there are that we don't want to send our data to China. Generally, you should probably want to keep your data to yourself or within your country regardless.

But I think the separate concern is, “Oh, the model itself—even if we host it in the US, is there a concern with the model itself?” To explain some of the concern there, I think the idea that some people have is—I don't know if you've seen, in those spy movies, where there's a code word and suddenly someone starts acting. You say the right word, and then they're in robot mode, where they're going to act adversarially.

I think the concern is that some of these models might secretly have that ingrained within them, where you say a trigger word and suddenly, even if it's hosted in the US, it's going to send data somewhere else or start intentionally trying to break whatever it is that you're doing. Suppose any nation were to try and do that.

Suppose they wanted to make a model that had one of these trigger words that was going to act adversarially. Theoretically, you would want to do that as late as possible, because if you do that in an early model and someone discovers it, they're literally never going to use your models ever again. I don't see that as a big concern.

Also, if you're deploying correctly—not as a consumer, but in the enterprise—data exfiltration or some of this adversarial stuff generally can be fought against. I do think, from a perspective of being quite patriotic, that it's pretty embarrassing that we don't have frontier open models in the United States. I do hope to see us reclaim superiority there.

Europe is significantly behind, especially on the model-development side. Do you think Europe is too far behind to catch up?

Probably on the frontier-model-lab side. There's so much to do on the infrastructure buildout and energy side of things. Again, the thing that's very difficult in the different parts of the world is that you have democratic countries where things generally are slower.

Suppose you say, “We need to do this thing.” You need to get a lot of support, convince certain people to do things, and pass legislation. It takes a long time. The benefit, though, is that theoretically we get this balancing act where we don't go too crazy in any one direction.

Other parts of the world that are more authoritarian are like, “This is the thing we're doing. We are doing it. We're acting now.” You get to move quickly. There's less correction, because what if you're going on the wrong course?

26. Data centres and the public backlash

In cases like AI, where it's pretty clear that, for the buildout, you need to build data centers and you need energy—and energy requires a huge amount of buildout as well, with a huge amount of lead time—you can act faster. In the West, things are slower, so that's one thing that goes against us. It's a little bit slower to get this stuff done, especially when there's all the politics that you have to deal with.

Do you worry about the public backlash to data-center development that we've seen? I think it's 40 out of 100 data centers post-approval that don't actually get built out in the end. Do you think data centers will be seen as a symbol of wealth concentration and technological superiority?

Yes, but I think that, at least in the United States, the beauty of having states is that we get some selection. We can have different experiments of what's it like for a state that says no to all data centers. Well, there won't be as many jobs created there, versus the states that do allow data centers to be created. People will prosper, have great jobs, and see the downstream benefits of it.

It's nice. It's like we have little petri dishes to test out and see how things work. That is the beauty of the United States. In Europe, it's tough. Europe had some good positioning a few years ago, a few decades ago, with nuclear power, which I think hasn't been delivered on as much lately. That would have been a world in which Europe had a way to bounce back a lot in AI on the energy side.

Well, 100%. I blame the Germans. And that's our German audience gone. Dude, I want to do a quick-fire round with you. I say a short statement, and you give me your immediate thoughts.

Nebius versus CoreWeave: who has a larger market cap in 5 years' time, and why?

To me—and this is speaking from a strongly biased application perspective—I'll take the grab bag. It doesn't matter. I actually hope for a world in which our users don't even know which one is under the hood.

For you, I would want CoreWeave to be bigger.

Why?

Because Nebius, I think, has more ambitious plans to be full-stack, which will eat into some of your plans in a way that CoreWeave won't.

Ambitions? What are ambitions? I don't think it makes sense for them to do that. Businesses need to think about their core competencies. If people are trying to expand beyond their core competencies, Kirkland & Ellis, great. Look, have fun. It's not your core competency. I don't think it makes sense.

Do we have a series of businesses like Nebius and CoreWeave where customer concentration is 90% of revenues? Will we see more of that?

Yeah.

Yeah, probably. Is that a bad thing or a good thing?

27. Selling without forward deployed engineers

It's bad if you're an investor in one of those companies, because it's a little riskier, but I think you can find a steady state. It's just scary. You know there's a sort of sword of Damocles above your head: if they ever leave, it's risky. It's risky.

Tell me, can you sell to enterprises today without an FTE model?

Yes. Have a good product. The thing about the FTE model that blows my mind is that, for us, when we do FTEs, the way I think about it is that their goal should be acceleration.

If there's a customer where, if we just give them our product, they'll scale to $1 million in 6 months, I'll throw in FTEs if they're going to scale to $1 million worth in 3 months. Great. We accelerated that. If I'm sending in FTEs as services, I'm not Accenture. I'm not trying to be Infosys or Cognizant or whatever. We are not a services company.

If we need FTEs to make the product work, we have a shit product. The point of FTEs should be to accelerate and get them consuming faster. If you're putting in FTEs because that's the only way you'll get a deal done, I'm sorry, my friend, you have a shit product.

28. Grindslop, sleep and treating teams like athletes

What do you think of the whole grind-slop element? We talked a little bit before the show with Nico at Corgi, which generated a little bit of discussion online.

Yeah, just a little bit.

Just a little bit of discussion online.

You're ruffling feathers, as always.

Dude, I said nothing. Honestly, it's like someone comes to your party and does something well. It's like, that's me. What do you think of the grind-slop?

I feel like a lot of the things we've talked about are things everyone needs to be wary of. Grind-slop comes from intermediate metrics. Generally, to do things, you need to spend time on them, so let's focus on how much time we spend instead of whether we're doing the thing right.

The analogy I use is: imagine trying to measure who won a basketball game by who sweated the most. You could sweat a ton, but look at the scoreboard. Are you doing what actually needs to be done or not?

For us, we want to focus on getting the best players. I don't care if you sweat a ton or if you sweat very little. If you're scoring a lot, great. We want you on our team. Generally, for most people, you have to sweat if you want to get things done.

But I think you're doing a bad job of hiring if you need to mandate certain crazy hours or need a bed in the office. Dude, get a good night's sleep. You don't need a bed in the office. Just get an apartment nearby that's nice and cozy and get 8 hours of sleep. If you, as an important member of your team at your company, can get your job done on 2 hours of sleep, you're not doing very high-leverage work.

But I did think it was an amazing opportunity for Eight Sleep to do an amazing social campaign. I would've delivered it. I would've got the founders outside being like, “We got you covered.”

It's funny. Wouldn't that be funny? When we were 30 people, we had what we called a surge—a pretty aggressive 2-week sprint. As part of it, I got everyone on the team Eight Sleep, fully free—whatever, $3,000 per person. The decadence of startups, right?

But I think the idea there is that we are optimizing for output. The people we're bringing onto the team are like SEAL Team Six, like the NBA All-Stars. It is worth every dollar to make them more productive and deliver on these ambitious goals that we have.

For me, I think Eight Sleep helps with my sleep. Great, let's do it. They're going to be better, have more of their wits about them, and be sharper. The type of engineering work that we do isn't just grunt work where the question is how we can spend as many hours doing it. We have droids for that.

The work that we do might require really deep thought—really, every ounce of brainpower that you have. If you didn't sleep well, you're not going to make as good a decision.

If I gave you unlimited money, what would you spend on today that you're not spending on?

I think, generally, we will see the best companies treat teams more and more like—whatever—SEAL Team Six or NBA professional athletes. Not in the way that Google did it, with, “Oh, you get a bounce castle,” and all this weird shit, but in the way that, with athletes, it seems like they're getting pampered, but it's kind of a burden.

Your diet is monitored. You have to do your hour-long massage after a game to make sure your muscles are recovered for the next game. You have to do an ice bath and all this stuff. It seems glamorous, but sometimes it's not.

I think spending on that type of stuff, but obviously in the more intellectual domain, is what more and more companies will do. If I could spend an incremental dollar to make every person sleep that much better, recover that much better, and be that much better at making decisions, it's probably worth it.

You're such an American. Do you know what I like? I like limoncello. You know what I like? I like smoking. Do you know what I want to do? I want to sit in the sun under the intense vitamin D rays and take in life with my friends. Yeah.

And you guys are optimized to the extreme. Did you see the Steven Bartlett video the other day? You might not know this guy, Steven Bartlett. He said, “I drank 2 glasses of wine and it ruined 3 days of my life.”

Yeah.

It was because I didn't sleep, and then the next day I ate more. I podcasted worse. I didn't go to the gym, and then I slept badly again. 3 days were ruined.

Okay, honestly, I get that, to be fair. The first year at Factory, I would drink a whiskey every night. And my argument was—and you'd probably agree with this—that for robustness, if you want to be a robust human, you can't have one drink that ruins the next 5 days of your life. The wind blows and then you're ruined, right?

So, to some degree, I get it. You want to have some of this stuff. But I also think, maybe, again looking to athletes, what they do is they have in-season and out-of-season. Maybe when you're in-season, you're locked in. You're not drinking. You're optimizing all this stuff with your Eight Sleep.

Then take a week off, go on the beach, drink some mojitos or whatever the hell people drink on the beach.

Well, out of season, you're Charlie Sheen.

Yeah.

29. Anthropic vs OpenAI

If you can recover after, to each their own.

Oh God, that would be the funniest thing ever.

Work hard, play hard. Yeah. Okay. You can invest in 1 company on IPO day. Sorry, dude. Anthropic or OpenAI?

In my mind, the answer here is, I think, they're approximately equivalent. To me, it doesn't really matter. The biggest reason that affects the EV is the volatility of the company. I think from a business perspective, they're both very well suited and kind of well positioned there.

So you were saying Anthropic?

30. Did Dario do AI a disservice?

Probably just the past is an indicator of the future, and there's just been more random, chaotic, turbulent events at OpenAI. But from a business perspective, to me, they're both great choices, but Anthropic.

Okay, we got that. Good. Can I ask: has Dario done a disservice to the ecosystem by saying, “We're going to take your jobs. We're going to take your jobs. We're going to take your jobs”?

Yes. Actually, this really upsets me. On the one hand, I maybe just implied Anthropic there, but on the other hand, I think that has been not only disingenuous and wrong, but it's really hurt the psychology of a lot of people—developers, just people in the world. It does AI a disservice and does the world a disservice, because this is, again, talking about the use cases, the problems that will be solved for society.

This feeds fuel to, “We should slow down AI. We should stop doing it.” And honestly, it's for selfish reasons that they did that, because if you're trying to raise unprecedented amounts of money—hundreds of billions of dollars, whatever—the best way to convince people to do that is to say, “All of capitalism is gone. The only company that's left will be me, so you better give us your dollars.”

Then suddenly, when it comes to an IPO, when all the humans and the people that you might be replacing now have money that you want them to put in your IPO, suddenly it's, “Whoa, whoa, whoa. Oh no, humans are pretty important. There are going to be jobs again. You know, we like you guys.” That pisses me off.

I totally agree. And what's ironic is the ones who've never said it are the ones who've never needed the money. When you look at a Zuck or a Demis, they've always had a very different stance from Sam and Dario when it comes to labor displacement and jobs.

Yes.

It's really interesting.

I mean, it's just a shame because, for all the philosophizing about AI and intelligence and all this stuff, incentives are driving the outcome, and the incentive is, “I want to raise a lot of money.”

Which legacy company do you think has most embraced AI?

Well, honestly, EY, the accounting firm, is one of our largest customers. I know you said that's shocking.

No, it's fucking shocking. You win.

They are so agent-native. It's crazy. They're one of our largest customers.

They just push it down the organization.

They're basically—they saw what happened with the cloud. They saw the scars of being late and not jumping onto it aggressively, and they have some great engineering leaders there who are like, “Look, this is going to be scary. Some people are going to get upset. It's not going to be the easiest thing, but we are going to make our org agent-native if it's the last thing we do.”

They were honestly pretty early to it as well. To me, I think one of the most interesting things is seeing that they're more agent-native than some startups, which is wild.

31. What he's changed his mind on

Brave new world. Final one: What have you changed your mind on most in the last 12 months?

What I've changed my mind on the most in the last 12 months is that there was a brief period of time where I thought it might be just 1 or 2 companies that ran away with being kind of the frontier and the best. What seems pretty clear to me is it's probably going to be at least 4 that are going to be approximately as good.

And that is a win. That is the win for humanity. The bad case for humanity is when there's 1 that's really, really good. I think there's probably going to be at least 4, if not many others. And that's something that it seems there's kind of growing evidence of, which, my sense is, is a hot take, because I think right now people are a little bit enamored with maybe 1 or 2.

Listen, Matt Damon, it's been so wonderful to have you on the show. I'm going to let you go back to Robin Williams. Kidding, dude. It's been so much fun. Thank you so much.

Thank you for having me.