软件与 AI 的未来｜Cognition 的 Scott Wu

I want to know why you describe yourself as salty. What does that mean?

I've just always been this way. As a kid, I hated losing. My first competitive memory ever was when I was in second grade. I went to a seventh-grade math competition. It was a middle-school competition held at the local university.

But you were 7.

Yeah. I was 7 or 8 years old. I was competing in the middle-school math competition. I took the math test, and then they were calling out the names: “Here’s who got third place. Here’s who got second.” I was waiting for my name to get called, but it wasn’t any of them. I remember being so pissed about that.

I can’t really give you a rational explanation for why it is.

It doesn’t have to be rational, but how much of your brain is dedicated to competition?

It’s all I do, honestly. I don’t know.

Oh, what do you mean, it’s all you do?

I think strategy games—I don’t know. The way even building a company feels the same. You’re calculating the moves. You’re thinking, “Okay, if you do this, then this happens, and then you do that,” and you’re calculating the different moves and what leads to success. It’s like a tree search, where you’re exploring the different options in the decision tree and trying to figure out how to lead to victory. That’s the only thing I do in my life.

So this is basically just—you don’t have memories from before you were like this, basically?

I think that’s right. I was a little brother growing up, and my older brother was 4 or 5 years older than me. Naturally, we’d play video games, and I would always be super salty there as well. It’s just always been like that.

I spent some time with Demis from DeepMind, and what’s interesting is that I draw a lot of similarities between the two of you because I’ve also spent some time with you. I was like, “Well, they’re both really smart, they’re both articulate, and they have a friendly UI, right?” But underneath that is this ruthlessly competitive drive. Demis—I think he said this publicly—said that half his brain is dedicated to competition, and a lot of that comes from his early days in chess.

Yeah.

What were you competing in when you were younger besides math competitions?

Basically everything. The main things were math and programming competitions. Ever since I was really young, that was my life. My whole goal was to become world champion of competitive programming. I did that all the time as a kid.

The really great thing about these competitions is that you compete in your school competition, and if you do well enough in that, then you qualify for the local or city competition. If you do well in that, then you get to the regional competition and the state competition, and then you get to go to the national thing, the international thing, right? It’s a very nice setup where, sooner or later, you get to meet people who are like you.

When I was a kid doing these competitions, that was all I really cared about. The people that I met through these national and international competitions were honestly more like my childhood friends than the people around me in Baton Rouge, Louisiana. We would hang out online, talk about math, talk about problems, and talk about all the other things, too.

I played basically all the different competitive games. I played a lot of Super Smash Bros. I used to go to tournaments for Super Smash Bros. It was a lot of fun. I played Melee, then Tetris. I played a lot of poker. I played some chess. I was okay at chess; I was not good. I played some Go.

My dad was a competitive Go player. My parents came to the U.S., in some sense, because of Go, which was a funny coincidence. My dad was in grad school in China, and he had a professor who really liked him. The reason he liked him was that my dad was a really good Go player. He was a 7-dan at Go, which, if you were to call it in chess, would be like, I don’t know, a 2,300 or 2,400 rating equivalent or something like that.

He would play with this professor on weekends, and my dad would generally win. They would talk about the games and stuff. Then that professor ended up moving to the U.S. to teach.

At the time, this was super early on in immigration from China to the U.S., so it wasn’t really a path that people knew they could take. The professor wrote my dad and said, “Hey, I came. It’s great. There’s so much more opportunity. It’s so much better. You should obviously come as well. I’ll help you with your visa application and help you apply to colleges here and everything.” My dad applied to grad school in the U.S., and that’s how we ended up here in the first place. I was born after we moved to the U.S.

So your dad was competitive in Go. What was your mom competitive in, though? I think I read that you said she might have been the most competitive person in your family.

Yeah. No, she was definitely the saltiest, for sure.

What does salty mean?

Salty just means that you take offense to the idea of losing.

Okay, I love that.

She would always be, “Oh, no, no, I’m better at this,” or, “I can beat you at this.” She played ping-pong a bunch growing up. Actually, she played on her school ping-pong team. She studied some amount of math and so on, but it was more her personality than any one thing that she put all of her competitive energy into.

I spent a lot of time reading the biographies of history’s greatest entrepreneurs.

Yeah.

I’ve always been fascinated by the fact that there are usually 2 different archetypes for the parents. On one side, you have Larry Ellison and Elon Musk. Their dads would literally tell them, “You’re worthless.” There are stories in Elon’s biographies where his dad would get in his face and yell at him for hours. Larry’s adopted father would tell him, “You’re never going to amount to anything.” They had this inner fire to disprove that, basically saying, “No, Dad, you’re wrong about this.”

Then you have the Sara Blakelys, whose uncle or even father is just like, “You’re really special. You have a lot of talents. If you put a lot of effort into this, you can do whatever you want.” Your mom falls more into the Estée Lauder-like category, where she would tell you, “These people are doing amazing things. You could do even better than them.” Correct?

I think they would have been happy enough if I had just gotten a more traditional, cushy job and done all of that. I don’t know that they specifically steered me toward entrepreneurship and being a founder.

But your mom gave you self-confidence.

Yeah, I think she always told me I was the best. [laughter] She was always extremely proud of that.

So you said she always told you that you were the best. Did she say that before there was evidence?

Yeah, I think so. Even when I was tiny, she would tell me that I was extremely talented. She was always a huge source of support for me and always believed in whatever I wanted to do. You would compete in math competitions, get these trophies, and growing up, we didn’t have pictures of our parents on the walls; we just had old math competition trophies. My mom was very intentional about it: “No, the thing that we value in this household—”

Other people’s trophies are the ones you want.

Ours—the ones that my brother and I won, of course.

We’re going to put up pictures of other people’s trophies, and you better damn sure replace them with your own. [laughter]

When I was pretty young, and my brother was as well, we both really liked these competitions and accumulated these trophies. Every time we got one, she would hang it up on the mantelpiece. I think what she really valued was education, and being the best was very important to her as well.

What matters to you more? Is the pain of losing—losing is worse than the thrill of winning? Because that's how you started it. You're just like, “I could not stand losing.” Let me back up. I was reading another Larry Ellison biography. He says this: “Listen, I'm addicted to winning, but I fear losing more than I love winning.” And I actually talked to Michael Dell about it, and he's like, “Yeah, it's just the fear of losing—the pain of losing—is way worse than the good feeling of winning.”

So, I think that's definitely true in terms of how it feels. In order to get anywhere, you have to lose a lot. If anything, a lot of the guys you're talking about—the way they got to where they are—is by losing a lot and having their share of wins along with that. You just have to put yourself out there and do a lot. So, it's kind of an interesting thing, to your point: I definitely feel the same way that losing feels way worse than winning feels good, but not by enough that it makes me want to stop trying, if that makes sense.

Yeah. No, I think it's impossible. I know your personality type; it's just impossible for you not to do this. Let's get into Devin. What does winning with Devin look like to you?

So, again, we're hypercompetitive, but the other thing about us is that we all had started our own companies before this. Our founding team was a pretty big founding team—it was 9 people—and most of us had already founded our own companies before. We had done different things. It was true for a lot of the early team, and we've always thought about this as the big one.

We want to go for it all. We want to be a generational business. We want to build a hyperscaler, and we want to go and do that. Maybe we'll succeed; maybe we won't. I don't know, but that's what we're going after. And I think, to me, what that means in our field, in building software, is that people sometimes use the term “coding agent” or “AI programming” or something.

I always hear that, and I think a little bit about, well, we're not always—I mean, we're not going to be interacting with code for that much longer. Our programs might not be the right level of abstraction, but I think what is always true, probably, is that it will be the human-computer interface. What I mean by that is that the way we think about Devin, if we're successful, is that Devin is the way humans can tell their computers what to do, because that was the whole point of software engineering anyway, right? It was just to be able to work with your computer and tell it what you want it to do. Doing that for the world is a massive opportunity, and that's what we're really excited to go after.

Why is that interesting to you? You just said we all did different things. We came together and we were like, “No, this is the big one.”

Yeah. I mean, the simple answer is we're a bunch of nerds. We're all programmers and built software and so on before, and now this is the idea of teaching AI how to do that.

Wait, say more about that—what you just said. We're all nerds. This is a big one because we can actually teach AI how to do that. We're all nerds who've spent the last 20 years of our lives just coding and making little things for the world. And the idea that you could teach AI how to go make things, and have everybody have an AI that can help them make things and feel the wonder of that—I mean, it's pretty sick.

Is that the idea—the single idea in the world that animates you the most, that gets you the most excited?

I think it is incredible. Yeah. Well, I would go further, too, and just say, at a zoomed-out level, I honestly think that—because everyone talks about AGI, everyone talks about, well, what is the future going to be like? What are human lives going to be once AI can do everything?

To some extent, I think the thing that is most human, obviously, is self-expression and creativity—having things that you want to make happen in the world and being able to go do those things, right? A lot of how I think about this is basically that we want to build the tool that gives everybody the power to go and make things that they want to make in the world.

My co-founder has a line that I've always loved: we've been spending all this time living in survival mode as a species, and now we're going to be living in creative mode. And I think that's right. Minecraft survival mode is where you're growing food, making sure you're safe from the monsters at night, and whatever. Creative mode is just like everything's up to you. You have all the resources at your disposal. If you want something to happen, it'll happen. The only question for you is what you want to make happen.

I think it's going to be amazing. And I think that is the world we're going towards, and we want to be the ones building that.

Okay, so talk about what Devin does today. And then I want you to flesh out that idea of where you see it in the future—it's just the interface between humans and computers.

Sure. Yeah. So, Devin is today what folks know as an AI software engineer. Basically, what that means is that Devin is a tool that anyone can use that will work with them end to end on building out software.

We work with a lot of the biggest companies across the world. We work with Goldman Sachs, Mercedes, a lot of areas of the US government at this point, and so on. We work with their software teams to help them build more and do much more. The thing about it is, in the last 20 or 30 years, “software is eating the world” is the famous line. I think it's very much true. It's still got a couple orders of magnitude to go, and in practice what it looks like is that teams use Devin to ship 10 times faster and to do 10 times more.

Okay, so that's where it's at today.

Yeah.

How do you get to where you're saying you might be just the interface between humans and computers?

So now we're going to get into a philosophical discussion of what it means to be a programmer or a software engineer, right? If you go all the way back, there was a time when programming was using the vacuum tubes—plugging all those in and having the machine do the arithmetic, right? The ENIAC was, in some sense, the first computer out there, although it was obviously very different.

Or it would have been filling out the punch cards, setting up or writing the assembly, or writing in BASIC or something, right? So we've gone through a lot of generations already, is my point. What does it look like going forward?

When you talk about programming, all it really comes down to is: how do you tell your computer what to do? Every single piece of software that you use—if you're using Instagram or TikTok or YouTube or whatever—that's a piece of software that somebody, or in these cases some pretty big teams of engineers, came together and thought through all these details: here's what I want it to do, here's how I want it to look, here's what I want this button to do, here's how I want to architect it.

Every single little decision was made by somebody, but the computer itself is then executing it according to what the wishes of its creators were, right? I think what we'll start to see is that abstraction will continue to climb. We kind of see this already. At this point, you don't need to know programming languages—you don't need to know Python or Java or something like that—in order to build your own software, right?

You can just say, “Hey, here's what I want. I want to make a cool website that does this, this, and that.” Or, for example, in my existing product, “Here's what we have today, and I want to change this thing or add this new plan or add this new feature,” and just have the agent go and do that for you, right? I think we're going to continue to go further down that axis.

One important distinction I'd make, which is what I think we'll see a lot more of in the near future, is that software today—the math only really works out to create software if it's going to be used at least a million times or something. Maybe it's 10,000 times or whatever. My point is, if you want to go build a product today, you need a whole team of engineers. Engineers are expensive. You've got to pay salaries.

You’ve got to go build all this out, and you’ve got to go and do that, right? You need that software to be used enough times—or to create enough value—for that to be worth it, right? Something like YouTube passes that test because obviously so many hours have been spent building YouTube, but way more hours have been spent using YouTube, and that’s what’s made that work out and made it feasible, right?

But there are so many things out there—very specific things—that only need to be used a few times or even only need to be used once, right? All of the white-collar work that we talk about today is very much like: all right, you wake up in the morning, and you’re going to look through these 15 LinkedIn profiles. You’re going to look for this and that, or you’re going to fill out these forms, or you’re going to do this data analysis and put this Excel sheet together with the research that you found. All of these things are things that could be done with software; it just obviously doesn’t make sense to hire a whole team of people to make you that piece of software, which you’re going to use 1 time and never again, versus just having the human go and do that themselves, right?

I think what we’re going to get to is a point where you’re just giving your instructions to that agent. On the back end, you don’t even have to look at this, but the agent is going to figure out, “Okay, here’s what I’m going to do. I’m going to write this code that’s going to go do this. I’m going to put a script in place that automates this part. I’m going to do this, and I’m going to do this.” That’s what’s going to allow it to actually go and do all these things.

But what you start to get to, as we’re saying, is that this is really just how you control your computer and how you do what you want to go do, right? You wake up in the morning and think, “Here’s what I want to go do.” You talk to your agent about it, and you figure out the task together. Once it understands it, it can do the literal part of, “All right, put pen to paper” on writing code. But that task is—you’re the one who’s deciding what to do.

So this ideal future that’s in your mind, how far away do you think we are from that?

Yeah. We’ve made a lot of steps toward it. I would say we still have a ways to go. You can use Devin today. You can use all the different coding tools today, and you can do a lot more than you could have done 10 years ago, or even 1 year ago or 6 months ago.

But it’s certainly not at the level that we’re talking about, where you’re neural-linked into the AI and can tell it exactly what you want it to do and what you want to see in the world, and just have the AI go and do that. When do we get there? It’s hard to say, but honestly, I think we’ll have solved most of that over the next 5 years or so.

In AI terms, 5 years is like a century in the rest of the world’s terms. Obviously, it’s crazy to imagine that things can change that much in 5 years, but I really think they will.

So this is related to something I heard you say, where you were saying that humans just have a really hard time understanding exponential curves.

So true. You see this in the progress itself. You see this in the scaling laws with the data. You see this in the revenue curves of the companies that are building AI—

Including your own.

And it’s very—you know, humans aren’t really wired for this, right? All of our inherent fight-or-flight responses, our ability to measure things—to vastly oversimplify, if you’re just fighting out there, foraging for food, or whatever it is, a good hunt will bring you a couple days’ worth of food. But obviously, with the kind of exponential curves that we deal with, the equivalent of a good hunt here could be 1,000 years’ worth of food, and we don’t have that intuitive signal in our brains to really understand that at a deep, native level.

People often underestimate, I think, how fast things can change and how fast the world can change. My parents even drilled this into me because they grew up in communist China. They came to the US, and even that—we’re talking about what were ultimately much slower scales of progress, in some sense, relative to what we’re seeing today—but even that was, as you can imagine, incredibly jarring to them.

You come to the US and everybody has a car. Everyone has all these household appliances and all these different things. It was a very different life when they grew up in China in the ’60s. They were much, much poorer, and it was very, very different, right?

Funnily enough, people got used to all these things pretty quickly, and now we can’t live without them. But I think we’ll undergo the same period with AI, where 5 years from now it’s going to be insane to think about all these things that we’re going to have. 10 years from now, we’re going to have forgotten that we ever lived without them.

What do you think you understand about AI that other people don’t?

The reason I asked the question is because we have some mutual friends. I would describe our mutual friends as some of the most AGI-pilled people that I know.

Yeah, and I feel every time I have a conversation with them, I’m like, “Even though I pay attention to this stuff, I feel like a toddler compared to somebody like you.” So I’m very curious: What do you understand about AI that most people don’t, including people within the tech industry?

No, you’re way too generous. Scott Wu

I don’t know that I have anything that interesting or that deep of an insight.

You say that because you’re used to it. Here’s what I’d say: The way that folks typically predict the future or think about what happens next is they pattern-match based on what they’ve seen historically. They say, “Okay, well, for 100 years it’s always been like this, and it’s probably safe to assume that it will be.”

And 99% of the time, that works great, right? In these particular periods where things actually move and there are real things that are different, those are the 1% of times where it truly is different. Rather than any kind of pattern matching, what really matters is thinking about things from first principles.

There’s the famous METR report, which was saying, a couple years ago, that AI would do about 10 to 20 seconds’ worth of human work without interruption. Then you’d have to guide it or direct it, or it would make a mistake or something like that, right? Ten seconds, 20 seconds—and that has just doubled every couple months, basically. Now we’re talking about hours of work.

So basically, an AI can take a task that would have taken humans hours of work. If you describe that task well enough to the AI, it will do the whole thing and come back to you with the result, and then you give it the next thing, the next thing, right?

If you ask, from a first-principles perspective, “Why can’t that be days, weeks, or months of work?” then what does the world look like if everybody has an agent that can just do months of work for them at a time? You get to a pretty different conclusion from what we’ve all seen and what we’ve all lived for the last several years.

I think that kind of first-principles thinking is, as different as it sounds and as crazy as it sounds, actually more correct than the simple pattern match, if that makes sense.

I think you’ve even taken this further, where you’re like, “What happens when they can work for a year unassisted?”

Yeah. I think that’s true, and I think we will get there. If Devin could work for a year without any human assistance—

Yeah.

What would you have it do right now?

Destroy your competitor.

All sorts of things. I mean, no. I still wake up and think about this in every different little thing that I run.

Dumb example: Yesterday, I was sending out a bulk email and trying to get the email formatter to work, and it’s kind of painful. Some of these things are still hard to use, or they don’t support a certain styling of the email that you want to use.

I think I had pasted something with indents, and it just couldn’t unindent them because the editor—I don’t know—there were some weird things where the editor would not allow you to unindent one part but not the other, or whatever. I was just thinking, “It’s really crazy that I’m still doing this,” basically, right?

In as much simplicity—or honestly, more simplicity—as it would take to explain this to another person, you could just say, “Here’s what I want to do. I’m just trying to make it look like this, and then this, and then that.” The rest of that execution should just be done for you, you know?

Then you get to the point where you really just get to spend all your time thinking about: What do you want to do? What do you want to build? What do you want to create? What are the things that you want to see in the world that aren’t there already?

But I think what makes it interesting about what you were saying earlier is that the more you increase the time, the more interesting it gets to me. There’s this guy named Edwin Land, who I won’t shut up about, and he was the founder of Polaroid.

Yeah. Steve Jobs’ hero. A lot of what we think of as Steve Jobs’ ideas literally just came from Edwin Land, down to the chairs and the table he would use for his presentations. It’s like the same thing that Land used in the ’70s at Polaroid.

And he thought of himself as a scientist, not an entrepreneur. Edwin Land died with, I think, the 3rd-most patents. He’s like Thomas Edison, some other dude, and then Edwin Land. And what he did is, he couldn’t figure out—he invented the industry of instant photography.

Yeah. Before, you took a picture and you’re like, “Hey, how’s it look?” “We’ll find out 2 weeks from now when we get back to Kodak.” I had no idea. And now, when you take a picture with a Polaroid, it’s like, “We’ll find out in 60 seconds when it dries.”

But that was black-and-white forever. When I read that part where you’re like, “Well, we’re going to have agents that can work on a problem for a year,” I didn’t think of what I would do, which is a question I just asked. I thought of Edwin Land hiring this guy. He’s like, “I want you to think about how we turn this from black-and-white into color.”

And before he could begin, the guy worked there and just thought for 2 years.

Sure. I’m like, that would be very convenient if I could have an agent attack this problem in the background while I’m working, because I can’t figure it out. They’re just thinking about how to attack a single problem for 2 years. In his case, this guy who solved instant color photography—it just took him 2 years of thinking to do it. So I’m going to push you on this a little bit more, because I don’t want your year-long agent to send bulk email.

Oh, I agree. Just to be clear, I think at some point you see this, right? When you’re talking about seconds, you’re literally talking about a specific command. When you’re talking about hours, you’re talking about giving it a task and having it do the task. And I agree with you: bulk email is not, for years, what you’re talking about. What you’re talking about is, you’re giving the agent a mission, basically, and it’s like—

That sounds way more fun.

Yeah, exactly. And this is like, what do I care about? I’ll give you a million different examples. The answer might be that there’s this one societal problem which is really important to me, and I think it’s a solvable problem. I think we can all be happy, but we really need to spread awareness about it. We need to get folks to understand the points of it, and we need to figure out how everybody should work together and coordinate on it. That’s a problem that an agent can think about, right?

Or even some of the sillier things, too. There’s this video game that I really like, but I wish—if I were making the game, here’s exactly how I would think about all these things. I feel like there’s this really cool idea if you could combine elements from this one game that I really like, but then incorporate some of the elements from this other game, and set the agent off on that mission: “Look, we’re going to go and make the coolest thing ever and the coolest game ever. We’re going to incorporate those elements. We’re going to think about how those nicely intertwine and work together,” right?

Or if it’s like, here’s this piece of novel science which I’m really passionate about. Materials have been created this way for years and years, but here’s this avenue of attack which I’ve been wondering about and thinking about: maybe there’s a different, novel construction of materials this way. Send your agent to go work on that for years—or months, right?—and have it study that, explore all these things, run its own experiments, and try all these things. I think all of these are soon going to be very possible. And to your point, it’s very different.

Yeah. I like the framing of sending them on missions.

Yeah.

I would have one that would pick the missions that I need to send other agents on.

Yeah, yeah, yeah. Then you’ll have the AI, which is the manager AI of the missions.

And I think we will continue—I always joke about how, if you think about our ancestors from hundreds or thousands of years ago, imagine them looking at us and what we do. You’re just pushing buttons, sitting in a room, and talking with other people, and you call that a meeting. That’s work for you guys? What do you mean, that’s work? I’m in the fields. I’m doing this every day. I’m going farming, making all of our clothes by hand, taking care of all these things, and that’s work. How can you guys call that work?

My point is just, I think what we will have going forward is going to look that different from what we have today. We will look at people who, as we say, just have these really interesting curiosities that they want to pursue. They have causes that they’re passionate about, fun ideas, or art that they want to create, and they’re sending off their agents in pursuit of those missions. They will think of that as work, and we will look at them and be like, “Wow, it’s kind of crazy that that’s what you get up and think about all day.”

When you started Cognition, did you know that you were going to try to make an automated software engineer? Was that your first idea?

I’ll say it was always 2 things. One, it was always related to code and software, which, again, probably has something to do with us all being programming nerds. And, 2, it was always around the idea that these would be real, multistep, iterative processes. That was, I would say, a real hot take at the time. This was 2023.

So this is when I met you guys. I was on a walk with a mutual friend of ours, and I remember first hearing this. You guys scaled from 1 million in revenue to, like, 500 million or something like that in, what, 20 months or something?

I remember, before you had any revenue, I heard the idea, and it was pitched as essentially an automated software engineer. I was like, “Holy shit, this is huge. You’re going after labor. This is a huge market.” And then you released—

Yeah.

I remember the demo video.

Yeah.

And then you got a lot of shit.

Yeah, yeah, yeah. A mix of, I think, the full polar ends of the spectrum, right? There were some people who were like, “This is the coolest thing ever,” and then there were some people who were like, “This is the worst thing that you could do.” Similarly, in terms of the capabilities, there were some people who were like, “Oh my God, everybody’s going to lose all their jobs tomorrow,” which is not what we’ve ever really believed or thought of this as. And there were also people who were like, “Dude, there’s no way this is ever going to work. This is totally a scam.”

So what was the criticism? Did you release the product earlier than you wanted to?

Well, look, it wasn’t even a product at the time, to be honest. It was more just a prototype or a demo of what was possible, and we had been working with it and playing with it for a few months at that point. We wanted to show people some examples of what it was capable of, because it was real—it was pretty insane for us to see as well. I remember the first time that it did a real task, I could not sleep that night. So it was just like showing that—

Say more about that.

The first task that Devin did was set up MongoDB for us. It’s a standard thing. It’s an issue that a lot of people run into all the time when they’re getting their initial database setup or whatever.

But we would just run into errors. You have this whole flow where you find an error, Google the error, find some message on Stack Overflow or whatever, or ask ChatGPT, and it tells you, “Okay, here’s one you should try.” You try that thing, run into a different error, paste that in, and then, for this one, at some point we were just like, “Okay, Devin, just try to go fix it. Just run the commands—do whatever you need to do.” And then it worked.

And you couldn’t sleep. It was truly—because, again, seeing the exponential curve ahead was incredible. This was a very specific case. It was the one success that we had, and it was definitely a way-better-than-average run.

But there was this feeling of, “Why shouldn’t all software and all products be built this way now?” You can just tell it what you want it to do and have it go do it. It’s kind of funny because, to your point, we did get hate in the beginning. We did not feel motivated at all by the hate to say, “Oh, yeah, maybe you guys are right. Maybe we should focus on the more chatbot Q&A-style product experiences.”

Maybe we should have done that more than we did. Maybe we should have done something in the middle. I don’t know. But I think for us, when we had seen that and when we had done these different things, all of the demos that we showed in our launch announcement were actual runs of Devin that we had done ourselves. We had run into them and been like, “Holy shit, this is insane for us.”

It was always kind of like, look, we can debate when, or what level of effectiveness, or whatever, but it’s just going to happen. That’s how we always felt about it.

So explain the process of iteration to go from that product, when you were getting a lot of hate. Actually, you know what? I called Jeremy Stern, who wrote this excellent profile of you in Colossus, which had a lot of parts that made me laugh my ass off, by the way. [laughter]

I was asking, “Tell me the stuff that didn’t make the profile.”

Yeah.

He said there wasn’t that much because you’re kind of an open book. A lot of people manage their media, and you have to talk about certain things off the record or whatever, but you were just saying everything.

He did say something about how you made the point that, when you released the first product, the benchmark was SWE-bench. What was the benchmark?

SWE-bench. It was 13%.

That was Devin. And your point was that that was already better than—

Yeah. At the time, the best known was 3% or 4% or something like that. But obviously, 13% still means you fail 87% of the time.

Was there a pronounced benefit from releasing early like that?

For sure. For us, by the way, if you think about the overall AI ecosystem, a simple way to put it is: We were late by a lot. OpenAI started in 2015. Google DeepMind and Google Brain are obviously more than a decade old already. Anthropic and other companies people talk about in the world had already been around for years. We were getting started in early 2024, more than a year after the ChatGPT launch, and a lot of the existing players were already there.

The same was true in code specifically. There was GitHub Copilot, which engineers had already been using for years, and that was very much the Q&A, autocomplete-style experience of working with AI.

When you start a company, you have nothing. You have no right to exist, is maybe one way to put it. It’s an interesting truth about the world. Obviously, startups succeed all the time, and the startup-versus-big-company story has been played out forever.

But there’s no reason, in terms of resources, people, or brand awareness, that you should win over what the big guys are doing. The reason you’re sometimes able to anyway is if you really plant your flag in the ground, put a stake in what you think the future is, and run like hell toward that.

If you turn out to be right on some of the core things—and I think we were wrong on a lot of things, to be clear—you can succeed. We were definitely early, which is a very fair criticism. A lot of the details and nuances, we learned and adjusted over time.

But the idea that you would work with AI as a co-worker rather than as a tool or a chatbot has obviously really grown over the last couple of years. I think it was very important for our brand, for recruiting, for customer work, and for everything else that we were the first ones to actually plant that flag in the ground.

So when did you have this idea that you guys were going to take a run, very ferociously, at these giant Fortune 500 companies—or even the U.S. Army, which uses you? Where did that strategy come from?

Funnily enough, that launch was in March 2024, and to your point, it went very viral. But again, we didn’t have any customers. We didn’t have any revenue. We didn’t really have—

I thought your first iteration of the business model was $500 a month or something, wasn’t it? Am I misremembering?

We had that, but that was actually later on. That was at the end of 2024.

Okay.

Initially, a bunch of people came and asked us for the product. We were like, “Guys, I don’t know if this is a product that’s ready for prime time, but if you really want to, you can try it and let us know.”

People ran with it, and we had to scramble to build a system where we could say, “Okay, let’s do a pilot or a POC.” We tried our best not to overpromise. We said, “Guys, it’s really early, but if you want to try it, you can try it.”

We did all these pilots, and perhaps unsurprisingly, they were all just failing. [laughter] That’s kind of natural. You could do some pretty cool things with it and some interesting toy demos or projects, but it certainly wasn’t ready to work on actual companies, real codebases, and so on.

From that point—this would have been April and May of 2024—we were making agents work with GPT-4. This was a very different era, so they were much more primitive agents. We kind of got to this question of, “Okay, what do we think it actually looks like when the agents start to get good enough for adoption?”

What we came to was that different tasks are different. There are some tasks that are really mundane and repetitive, where you’re doing the same tedious thing over and over again. There are other tasks on the opposite end of the spectrum that are really tough, like architecture problems or deep issues where you have to understand all the context and have all the know-how to know how to fix them.

People were trying to use Devin for all of those, and we were failing at it, understandably. The question was, “What are the natural tasks? If there’s a first task that’s going to have product-market fit and real value from these agent experiences, where it can do the whole thing end-to-end, what is that going to be?”

We said, “It should be some of these really repetitive, tedious tasks.” It’s not so cut-and-dried that you can have an automated script do the exact same thing every time. It does take some intelligence and some amount of meandering, but it’s repetitive enough, scoped enough, and on a tight enough feedback loop that you can have an agent do it. It would be able to diagnose and fix the problem.

That brought us naturally to some of the initial use cases, which were things like migrations and version upgrades—helping people upgrade from Java 7 to Java 8, for example. As you can imagine, enterprises had these massive codebases. You might have a 50,000-file codebase where you have to make the same 8 changes in each file. You have to be a little thoughtful about how you make the trade-offs, but it’s a very repetitive task.

Our first success ended up being with a company called Nubank, the biggest bank in Brazil by market cap at the time. The use case was one of these big migrations, and we had a custom Devin that was extremely optimized for doing that.

As we grew from there, later on, as things got a little more mature, we had both a self-serve business and an enterprise business, and so on.

But I think from the beginning, we had always seen that building software in the real world and managing massive products that millions of people use every day was substantially different from just building a cute website or a cute demo from scratch. I think we really leaned into that.

Naturally, all the Fortune 500 companies—or all the biggest companies in the world—are software companies in 2026. Even the ones that folks don’t necessarily think of that way, like Walmart, CVS, JPMorgan, and Mercedes, are all software companies. They have massive teams of software engineers, and they have tons of things that they’re building, shipping, and maintaining.

That became a natural thing for us to work on because we had learned very early on that we wanted to work on real problems and things that matter—things that people actually care about.

So wait, what percentage of your revenue is coming from enterprise?

Today, around 75–80%.

What are people using on self-serve?

Yeah. We have a lot of teams that use it through self-serve. That has grown quite a bit lately as well. We have startups like Exa, which uses it a ton, OpenRouter, and Bilt.

I ran into someone in my apartment elevator the other day. They said, “Oh, you’re the Devin guy. We use Devin.”

Are you Devin?

I’ve gotten that as well. “Hey, are you Devin?” I was like, “I’m actually not, but that’s okay.”

Why don’t we call it Scott?

Okay. Can you walk us through this? I’m very curious. Let’s say I’m a big enterprise. I contact you. Walk us through what happens when I become a customer.

Yeah. It starts with education. Everyone has gone crazy over AI and agents, and those have obviously been the buzzwords of the last 6 months. People want to know more about this, but there’s still a lot of detail involved in what it actually looks like to deploy them.

We show them what this looks like. We talk them through how we work with teams and how we partner with them, how we direct them to the right use cases, how we give them guidance on maximizing their ROI, and which projects are or are not feasible with agents.

From there, enterprises typically have very messy processes. For most software vendors—or, more generally, vendors they want to work with—adopting software and giving it access to all their repositories, getting through security, and everything else can be, as you can imagine, a 12- to 18-month cycle for a massive bank. What we naturally do is work with them to figure out how to do that as fast as humanly possible.

Did you send employees down to South America for Nubank?

For Nubank, honestly, our entire team was the forward-deployed team. We all flew to Brazil. It was the first case, and let’s be real: agents did not generally work at that point. A lot of it was figuring out how to make it work in a very specific context.

We were all there, sitting with their engineers and understanding, “This is what you do in that case. This is what you do in that case. This is what Devin needs to know, and Devin needs to be able to read these things.” We were debugging their exact problems.

Obviously, it’s not like that at all now. We’ll get to where it is today, but the idea was not that we deployed a team. We deployed the whole company.

Oh, yeah. It was getting the first customer, obviously. I wonder what they thought of that.

It was literally, “Let’s go through all these different things that you think could make sense. Let’s go through each one. Let’s try some of it manually ourselves to understand what the task looks like. Let’s see if we can teach Devin to do this correctly and build the right orchestration for Devin to be able to do it.”

Building the product was almost like building it for one company. It was fun.

So what do you do today?

Today, it’s much more self-serve. Agents are so much more capable, obviously, and we’ve figured out a lot more things with the onboarding experience.

A lot of it is that we’re saying these cycles typically take 12 to 18 months, and we try to get deployed with customers within 3 months. A lot of that requires them to first appreciate that it’s a priority. If you have 25,000 software engineers in an organization that costs $10 billion a year or something, and you think you can move 3 times faster, usually it is a priority.

Then it’s figuring out how we get through the security reviews. We can deploy in their private cloud. We have very strict data agreements and tight firewalls around all these things.

Do employees physically have to go? Are these deals so big that your employees are working with a specific company, and only that company, for a period of time?

Typically, no. We have a full forward-deployed motion, but a lot of what that looks like is user education and guidance. We still fly out to see customers and work with them, point them to the right use cases, teach them how to be successful, help them with their setup and their playbooks, and do all these things for using Devin.

But it’s much more, “Look, we’re here to assist you. We’re giving you a lot of this direction, but you and your team are the ones using Devin and running everything,” right? That’s a lot of what it looks like.

We’re set up to be as incentive-aligned with our customers as possible. A lot of it is not just, “Here’s this tool. Hand it over to your engineers and find out what they say about it.” It’s, “Let’s figure out what initiatives you really care about, and we’ll point you to those initiatives. Here are the projects where we think Devin can make you 10 times faster. We’ll tell you which ones it can’t, what workflows you should use instead, and how you should get to value instead.”

It looks more like that than literally using Devin on your behalf.

Wait, tell me more about how you align the incentives between your company and your customers.

A couple of things here are really important. One is just working with them to clearly define the ROI. I think that’s very important. Everyone has been talking about this over the last few months: People are going crazy on their token budgets, and one engineer can spend so much. You want to know that it’s actually creating real value for you, and you want to be able to identify where you are.

You think all this stuff is a little crazy.

I think it is directionally correct, but there are definitely some places where people have gotten carried away. People talk about, “Yeah, we rank our engineers by how many tokens they’re spending.” Let’s try to rank people by how much output they’re actually producing, or how much good work is actually getting done. Obviously, the math works out: the GPUs are expensive, but if your engineers are actually able to ship 3 times more, then it’s very clearly worth it. You just want to make sure you’re doing it the right way, right?

Obviously, a lot of that is about tying it to specific outcomes: What are the actual tickets that are getting done? What are the projects and initiatives? This project was scoped out for 18 months, was going to be handed off to an outsourced contractor, and was going to cost you $15 million. Let’s talk about how you do this all internally with your own team, do it for $1 million, and get it all done in 3 months—stuff like that, which I think is super important.

The second thing, which I’ll call out especially, is just being neutral. A lot of discussion, obviously—what do you mean by that? Being neutral with respect to all the labs and the models themselves.

I think you said you like being Switzerland.

Yeah, we like being Switzerland. Exactly. I think it’s an important thing: We are just as incentivized as they are to figure out how to make their token spend efficient. Devin is purposefully meant to be a compound AI system. All of these different things can be used for the right task, or even the right subtask of a task.

Describe what a compound AI system is.

Yeah. For each different part of your use case, imagine you tell Devin, “Hey, a customer just reported this bug. There’s a Jira ticket. Can you take a look at the ticket and go solve the whole thing?”

What does Devin actually go do? First of all, Devin is probably going to investigate and understand what the report says and what’s going on. Second, as any engineer should, it’s going to try to reproduce the bug itself. It’ll say, “Okay, let me spin up the product locally. We’ll click around, try to follow the same steps that they followed, and see if I can make the bug happen as well.”

If you did that, then you’re looking at the logs. You’re trying to figure out what went wrong. You’re pinpointing the particular files or the flow of commands that led to this. Then you do the debugging, go and test it again, make sure everything works, and put it up for review.

It turns out that there are different models that are good for different parts of these tasks. Across different tasks, obviously, some are really crazy, hard ones where you want to use max thinking and the very best models you can get your hands on in the world. Many other tasks are boilerplate or repetitive enough that what you care about is getting them done really fast, getting an immediate answer, having the ability to verify that it was correct, and then making sure it was as cheap, fast, and efficient as possible.

Rather than being pegged to 1 model and saying, “We’re only going to serve you GPT for this,” or, “We’re only going to serve you this,” Devin can use any of the different models it has in its arsenal, including models from Anthropic, OpenAI, Google, and others. It can also use our own models or open-source models out there, and it will dynamically choose the right models for these tasks.

How do you think about this? You’re a customer of theirs, but also a competitor.

No, I mean, look, I think in practice there’s a lot of positive-sum work for us to do together. The way that we think about ourselves is a couple of things. One, software is the only thing that we care about, obviously. There’s a lot of value in focus, in building products, and in building our whole motion specifically around that.

Say more about how you think about focus and the value in it.

Back to what we were saying about startups: Why do startups ever win at all? If you do everything, you will lose to Microsoft or Google, which does everything but also has trillions of dollars more resources, 100,000 more people than you, and infinitely more brand name.

The way that you build a real, lasting business or product is by really, really focusing and narrowing on 1 specific thing, making a very concentrated bet, and then, obviously, your bet has to end up being right. From there, it’s a lot of really tight execution.

In software, I love the quote from Daniel Ek at Spotify. People were saying, “YouTube is trying to do this, and Apple has Apple Music. Why should there be a Spotify?” He was like, “I can give you all the other reasons, but the truth is, we’re just going to care way more about music than they are.”

I think for us it’s true. We are just going to care so much more about what it looks like to build software end to end at Goldman Sachs or at Mercedes-Benz, or something like that. There’s a lot of nuance and messiness in that. It’s not as simple as, “Here’s a sandbox algorithms problem. Go and code me the correct 30-line program that solves this problem.”

It’s about how you work with all the messiness of the real world. How do you understand the codebase as it exists today? How do you collaborate with all the humans working on it? How do you plug into their ticketing system? How do you give the agent the ability to test its own code and run everything locally? All of these are obviously super hairy, messy problems, and that’s what we care a lot about.

From that perspective, it’s a very nice relationship. The labs have their own products, and I’m sure they’ll continue to do more, but in practice there are a lot of nice ways for us to cooperate. Being Switzerland means that folks can work with us and trust that we will route them to the right models, optimize the price-performance for them, direct them to the right use cases, and so on. We’re not incentivized for them to spend more on the models either. We’re incentivized for them to get value and output from them.

I love that you use the example of the war between Spotify and the rest of the industry, like Apple Music. Jimmy Iovine, who also came on the show and is now actually a friend of mine, called me yesterday about this, because anytime anything happens with AI music and Spotify, he calls me.

We talked about this because he was like, “You don’t understand.” He had 40 years of experience in the music industry and all the relationships. He was like, “Apple bought me for $3 billion. Spotify, at the time, said, ‘We’re going to go head-to-head with them.’” Spotify only had 3 million paid subscribers at the time. Jimmy was going to fight the war.

He talked about how that wound up being huge. He thought it was an asset; it was a huge liability. Yeah.

Because they’re like, “We don’t give a shit about getting a couple million more, tens of millions or more, of subscribers. We invented the most successful consumer product of all time.”

And he said something like, “They just clipped his wings.” He tried to do something like, “Oh, this isn’t important to us,” where Daniel was going to die if he was not successful. He just cared about it a lot more.

Dude, even in the 2 years that we’ve been around, I’ve heard so many different versions of the same argument. When we started, as you can imagine, this was true for us. It was true for everybody in the space building the space.

The number-one pushback that you would always hear from investors and other people was, “Microsoft already has GitHub Copilot. Isn’t everybody just going to use that?” It was a very reasonable thing to say in some sense, because they did own all of GitHub, they had the partnership with OpenAI, and they had VS Code.

I think in practice, the reality is that there’s so much more innovation and so much more to build. Microsoft is a great partner of ours, and we do a lot of things together and build even more together.

People have said this forever: “Why should startups exist?” You can give all the rational reasons. It was like, “Google could do this,” or, “Facebook could do this.” Why should Datadog exist? Why should Snowflake or Databricks exist, or whatever? The clouds care about observability too. The clouds care, and the reality is that it’s, in some sense, a bit of an uncreative way to think about things.

10 years ago, it was like, “Google could do this,” or, “Facebook could do this.” It’s like, why should Datadog exist? Why should Snowflake or Databricks exist, or whatever? The clouds care about observability too. The clouds care, and the reality is that it’s, in some sense, a bit of an uncreative way to think about things.

I think if there was one thing—that there was some end-state future that we all knew was coming, where everybody was just working toward it and whoever had the most resources toward it would win—then, of course, we know who has the most resources today, right? But if there are millions of problems out there to solve, there are lots of different things to work on. The world is dynamic; things change all the time.

There are lots of new ideas, opportunities, and ways to build new products. The reality is that there are lots of different niches to own. There are different things to bet on and different focuses to spend your time on. I think that will continue.

For better or worse, after the last couple of months of news, Cognition has become a bit more known as the folks betting on independence, in some sense, because obviously there have been some high-profile acquisitions.

It's funny that you said that, because one of the questions I wanted to ask you—which you probably won't answer—is how many different acquisition offers you've had.

I will not answer that question.

Dozens more?

Dozens is a lot, dude. I don't know about dozens, but okay.

At the same time, it's probably only a handful that could actually afford to buy you if you would sell. But the number of times they keep coming back—

Oh, I see. Well, yeah. Anyway.

This independence part is really interesting to me. We were on the phone—you, me, and a mutual friend on a three-way call—I think it was actually last summer. We won't say the company, but I was like, “Scott, what do you think about an acquisition with X? Not X, the platform—X, the [blank] company.” You were like—

“I don't know how we'd be able to afford them.”

You assumed I meant you buying them, and I just fucking cracked up laughing. I was like, “That's hilarious.”

No. I think there are folks taking acquisitions and doing things. There are some big, high-profile ones, and I think those are great, to be clear. It's a very reasonable path for an exit.

For us, we've all come into this wanting to build a generational business. That's something we're really excited about. Today, I've seen more of this nihilism where people say, “Maybe it's too late. Maybe it's not possible anymore.”

What does that mean? What's not possible? Maybe it's not possible to build a new independent business because everything else—

The labs are going to do everything.

All the opportunity is taken, you know, and it's like, guys—

Those people aren't founders. Founders are rationally optimistic.

Yeah.

They're just not founders.

Yeah.

They believe, even when there's no evidence that they should succeed, that they will succeed.

So people like that just need to get a job. I'd say this even with Devin: If you think that going onto the Devin web app—or any other coding product out there today—and giving it instructions the way that you do right now, getting the pull request out, reviewing it that way, and building software that way is going to be how software is built forever, then nothing will change. But I think we have 10 more generations of these different product experiences to come. Building those and creating them—that is what innovation is. That's what's going to happen.

The advantage I have from reading all these biographies of entrepreneurs for the last decade is that you're studying somebody's life. In many cases, if they were so successful that a book was written about their life, you almost get a minor in new-industry creation every single time. Every time, it's just like, “We're too late. It's over. There's no more opportunity.”

The best definition of a business I ever heard actually came from Richard Branson, who said, “All business is an idea that makes somebody else's life better.” If that's true, which I believe it is, then there is infinite opportunity now and in the future, because there are infinite ways to make somebody else's life better. That's all business is.

So this idea that we've reached the end of history is just bullshit.

Yeah.

I don't mean to push you on this. I'm genuinely curious, though, because I know you're already rich. I don't think money is your north star, based on the conversations we've had in the past, but there has to be some crazy number somebody could throw at you where you'd be like, “Fuck, I have to take this.”

Not really. People have asked me before, “Okay, but really, would you guys sell?” This is what I'm doing right now. The way that I say it sometimes is that we would sell if we thought it was the most ambitious thing to do. It's kind of an oxymoron, obviously, but that's my genuine answer, in the sense that that's what we care about.

It's funny that you talk about money. I don't even have a car.

You live in an apartment. I just realized—

I rent an apartment. I don't know. I think I like eating sushi. That's fun. Sushi doesn't cost that much; you can do that off an engineer's salary as well.

Just so you know, if this is true, then you are the type of entrepreneur I find the most fascinating in the world. When startup founders come talk to me, I don't give a fuck about their startup. I ask the same question: “Is this your last business?”

Yeah.

Right. I asked Kareem from Ramp that before we did this deep partnership. It's like, okay, you could sell it—the Zuckerberg example. People ask, “Why didn't you take a billion dollars?” You own 25% of the company; you would have made $250 million. And he's 22 years old. He's like, “What would I do with the money? I would just start another social network. I kind of like the one I have. I just want to build shit anyway, so what do I do?”

The other element of this, which I think is tied into you, is that I feel like you're having a lot of fun. Even being around here, your company is weirder in the composition of its people because it's literally just all nerds, and I think you like it. You know what I mean? Usually there's a mixture of the nerds and some other people, but you've built this almost like your social network is—

I'm not here if you're not a nerd.

Yeah. It's a physical manifestation of your social network. The way I would describe it is like—

My friends are all also nerds in a way.

I love that. The people I'm most interested in are the ones where there's no price. The example I use is: Go ask Steve Jobs, “I'll give you $2 trillion, but you can't work on Apple.” What the fuck do you think he would have said?

Yeah, I agree.

No, he's like, “Well, I want to work on Apple. That's what I want to do.”

There's a funny thing I wanted to tie together, which I didn't find the opportunity to mention until now: You're currently the 2nd-wealthiest entrepreneur from Baton Rouge, Louisiana.

I don't know if you remember, but we've talked about this.

We talked about this, yeah, yeah.

Yeah.

The first one is Todd Graves, who owns Raising Cane's. I talked to him; he was on the show.

Chicken fingers are a great business, but I'm a customer of that business as well.

You tell him you'll give him $100 billion for his chicken-finger empire.

Yeah. He's turned down—and I know this for a fact; I've talked about it—crazy acquisition offers. He's like, “I don't care about it. It's not the money.”

Yeah.

For us, again, this isn't rational, but the way I would describe it is that I feel—and I think all of us do—that it would be one thing if we tried, gave it our all, and just weren't good enough. That would be an outcome I could live with. I'd be salty as hell, but it would be an outcome.

If we felt like we could have gone for it all, pushed harder, and we didn't, I don't think we would live with ourselves. That's the answer. Why are we so excited to do this? Why do we spend all this time on it? We want to achieve our potential and build what we were meant to build. Maybe that's something; maybe that's nothing. But you'd rather find out and see.

Yeah. I think this idea of “you have one life, go” is a perfect place to end. Thanks for the time, Scott.

Yeah. Thanks for having me.