ソフトウェアとAIの未来——Cognition / Devin のスコット・ウー

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

Yeah. I was like seven or eight years old. I was competing in middle school math, did the math test, and they were calling out third place, second place. I was waiting for my name, and none of them were me. I just remember being so pissed about that.

I mean it's all I do honestly. Strategy games—I don't know. Even building a company feels the same. You're calculating the moves. If you do this and then this happens and then you do that—you're calculating what leads to success. It's like a tree search, exploring options in the decision tree, trying to figure out how to lead to victory. That's basically the only thing I do in my life.

I think that's right. I was a little brother growing up. My older brother was four or five years older, we'd play video games, and I'd always be super salty there as well.

Basically everything. The main thing was 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. School competition, then city, regional, state, national, international—sooner or later you meet people like you. Those people I met through national and international competitions were more my childhood friends than people around me in Baton Rouge, Louisiana. We'd hang out online, talk about math, problems, and everything else. I played basically all the different competitive games. A lot of Super Smash Brothers—I used to go to tournaments, played Melee. Tetris, a lot of poker, some chess—I was okay, not good. Some Go. My dad was a competitive Go player. My parents came to the US in some sense because of Go. Funny coincidence. My dad was in grad school in China and had a professor who really liked him because dad was a really good Go player—like a seven dan, chess equivalent maybe 2300–2400 rating. He'd play the professor on weekends, generally win, talk about the games. That professor moved to the US to teach. Super early for immigration from China to the US—not a path people really knew. The professor wrote dad: hey I came, it's great, so much more opportunity, you should come, I'll help with visa and college applications. Dad applied to grad school in the US and that's how we ended up here. I was born after we moved to the US.

Yeah. She was definitely the most salty I would say for sure.

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

She would always be, oh no I'm better at this, I can beat you at this. She played ping pong on her school team. She studied some math. But it was more her personality than any one thing she poured competitive energy into.

I think they would have been happy enough if I just got a more traditional cushy job. I don't know that they specifically steered me towards entrepreneurship and being a founder. But—

Yeah. She always told me that I was the best, and she was always extremely proud of that. When we would go to these math competitions—

Yeah, I think so. Even when I was tiny, she would tell me I was extremely talented. Huge source of support. Always believed in whatever I wanted to do. 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 what we value in this household.

Ours—ones that me and my brother won. Every time we got one she'd hang it up on the mantelpiece. Education was really important to her, and being the best was very important to her as well.

That's definitely true in how it feels. To get anywhere you have to lose a lot. A lot of those guys got where they are by losing a lot and having their share of wins—you have to put yourself out there and do a lot. Losing feels way worse than winning feels good, but not by enough that it makes me want to stop trying.

We're hyper competitive, but also we all had kind of started our own companies before this. Founding team was pretty big—nine people—and most of us had already founded our own companies. True for a lot of the early team. We've always thought this is the big one. We want to be a generational business, build a hyperscaler. Maybe we'll succeed, maybe we won't—but that's what we're going after. In building software, people use terms like coding agent or AI programming. I always think we're not going to be interacting with code for that much longer—our programs might not be the right level of abstraction. But what is always true is the human-computer interface. The way we think about Devin if we're successful: Devin is the way humans can tell their computers what to do. That was the whole point of software engineering—work with your computer and tell it what you want. Doing that for the world is a massive opportunity. That's what we're really excited to go after.

The simple answer is we're a bunch of nerds who were all programmers and built software. Now the idea is teaching AI how to go do that.

We're all nerds who've spent the last 20 years coding and making little things for the world. The idea that you could teach AI how to make things, and have everybody have an AI that can help them make things and feel the wonder of that—it's pretty sick.

I think it is incredible. At a zoomed-out level, everyone talks about AGI and what human lives will be once AI can do everything. The most human thing is self-expression, creativity—having things you want to make happen and being able to go do them. We want to build the tool that gives everybody the power to make things they want to make in the world. My co-founder has a line I've always loved: we've been living in survival mode as a species, and now we're going to be living in creative mode. Minecraft survival mode—growing food, safe from monsters at night. Creative mode—everything's up to you, all resources at disposal. If you want something to happen, it'll happen; the only question is what you want to make happen. That's the world we're going towards, and we want to be the ones building that.

Devin today is what folks know as an AI software engineer. A tool anyone can use that works with them end to end on building software. We work with some of the biggest companies—Goldman Sachs, Mercedes, a lot of areas of the US government—and their software teams to help them build more and do much more. Software is eating the world is still true; still got a couple orders of magnitude to go. In practice teams use Devin to ship 10 times faster and do 10 times more.

Now a philosophical discussion of what it means to be a programmer. There was vacuum tubes, ENIAC, punch cards, assembly, BASIC—we've gone through generations. Programming comes down to how you tell your computer what to do. Every app—Instagram, TikTok, YouTube—teams thought through every detail and the computer executes the creators' wishes. Abstraction will continue to climb. You don't need Python or Java to build software—say what you want, change a feature, have the agent do it. We'll go further down that axis. Important distinction: software today only works out if it'll be used at least like a million times—or 10,000. You need a team of expensive engineers. YouTube passes that test—way more hours spent using it than building it. But so many things only need to be used a few times or once. White-collar work: wake up, look through 15 LinkedIn profiles, fill forms, data analysis and Excel with research—all could be software. Doesn't make sense to hire a team for software you use once versus doing it yourself. We'll get to giving instructions to the agent. On the backend the agent figures out code to write, scripts to automate—you don't even have to look. It's how you control your computer and do what you want. Wake up, talk to your agent about what you want, figure out the task together. It does the pen-to-paper of writing code; you're deciding what to do.

We've made a lot of steps. Still a ways to go. You can use Devin and coding tools today and do a lot more than 10 years ago—or even one year or 6 months ago. Not at the level of neural-linked into the AI telling it exactly what you want. When? Hard to say, but I think we'll have solved most of that over the next 5 years. In AI terms 5 years is like a century. Crazy to imagine that much change in 5 years, but I really think it will.

So true. Progress itself, scaling laws with data, revenue curves of AI companies—including our own. Humans aren't wired for this. Fight-or-flight, measuring things—a good hunt brings a couple days of food. With exponential curves, a good hunt could be a thousand years of food. We don't have that intuitive signal. People often underestimate how fast things and the world can change. My parents drilled this into me. They grew up in communist China, came to the US. Even slower scales of progress relative to today were incredibly jarring—everybody has a car, household appliances—very different from the much poorer 1960s China they grew up in. People got used to it quickly; now we can't live without them. Same period with AI: 5 years from now insane what we'll have; 10 years from now we'll have forgotten we ever lived without them.

You're way too generous. I don't know that I have anything that interesting or that deep of an insight.

There's the famous METR report—a couple years ago AI would do about 10 to 20 seconds of human work without interruption, then you'd guide it or it'd mistake. That's just doubled every couple months; now we're talking hours of work. Describe a task well enough, AI does the whole thing and comes back, then the next. First principles: why can't that be days or weeks or months? What does the world look like if everybody has an agent that can do months of work at a time? Pretty different conclusion from what we've lived the last several years. That kind of first principles thinking—as crazy as it sounds—is one of those times where it's actually more correct than the simple pattern match.

Yeah. I think that's true and I think we will get there.

Destroy your competitor—all sorts of things. No—I still wake up thinking about every little thing. Dumb example: yesterday sending bulk email, trying to get the formatter to work—painful. Pasted with indents, couldn't unindent because of weird editor constraints. Crazy that I'm still doing this. As simple as explaining to another person what I want—execution should be done for you. Get to spending all your time thinking what you want to do, build, create—things you want to see in the world that aren't there.

I agree. When you're talking seconds, a specific command. Hours—give a task and have it do the task. For years, you're giving the agent a mission. Way more fun. A million examples. A societal problem that matters—spread awareness, get people to understand, figure out how everyone should coordinate. A video game you love—combine elements from games into the coolest thing ever. Novel science you're passionate about—materials constructed differently—send the agent to work months or years, study, explore, run experiments. All soon very possible. Very different.

Then you'll have the manager AI of the missions. Imagine ancestors hundreds or thousands of years ago looking at us: you're just pushing buttons, sitting in a room talking, calling that a meeting—that's work for you guys? I'm in the fields farming every day. What we have going forward will look that different. People with interesting curiosities, causes, fun ideas or art—sending agents on those missions—will think of that as work. We'll look at them and think it's crazy that's what they get up thinking about all day.

It was always two things. One, always related to code and software—probably because we're all programming nerds. Two, around the idea that these would be real multi-step iterative processes. A real hot take at the time—this was 2023.

Yeah.

Full polar ends of the spectrum. Some: coolest thing ever. Some: worst thing. On capabilities: oh my god everybody's going to lose their jobs tomorrow—not what we ever believed—and also, no way this ever works, totally a scam.

It wasn't even a product at the time—more a prototype or demo of what was possible. We'd been working with it a few months and wanted to show examples of what it was capable of. Pretty insane for us to see as well. First time it did a real task, I could not sleep that night.

The first task Devin did was set up MongoDB for us. Standard issues people run into getting initial DB setup. You find an error, Google it, Stack Overflow or ChatGPT, try that, different error—loop. At some point we said okay Devin just try to fix it, run the commands, do whatever you need—and then it worked.

Truly—seeing the exponential curve ahead. Very specific case, better-than-average run. Feeling of why shouldn't all software and products be built this way now—tell it what you want and have it do it. We got hate in the beginning. Not motivated at all to pivot to chatbot Q&A style. Demos in the launch announcement were actual runs of Devin we'd done ourselves and thought holy— this is insane. We can debate when or effectiveness, but it's going to happen. That's how we always felt.

SWE-bench. Yeah. It was like 13%.

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

For sure. In the overall AI ecosystem we were late by a lot. OpenAI started end of 2015; Google DeepMind/Brain more than a decade old; Anthropic and others around for years. We were getting started early 2024—year-plus from ChatGPT launch. Existing players already there. Same in code—GitHub Copilot engineers had used for years, Q&A autocomplete style. When you start a company you have nothing—no right to exist. No resources, people, brand awareness the big guys have. No right to win. Sometimes you win anyway if you plant your flag in what you think the future is and run like hell—and if you're right on core things. We were wrong on a lot of details and nuances; learned and adjusted. Early is a fair criticism. But working with AI as a co-worker rather than tool or chatbot has really grown. Important for brand, recruiting, customer work—to be first to plant that flag.

Funnily enough, that launch was March 2024, went very viral, but we didn't have any customers, any revenue.

That was later—end of 2024. Initially people asked for the product. We said not sure it's ready for prime time, but if you really want you can try. Scrambled to build pilot/PoC systems, tried not to overpromise. Unsurprisingly they were all failing. Cool toy demos, not ready for actual company codebases. April–May 2024, making agents work with GPT-4—much more primitive era. We got to: what does it look like when agents get good enough for adoption? Different tasks differ—mundane repetitive vs tough architecture. People tried Devin for all and failed. First task with PMF and end-to-end value: repetitive tedious ones, not so cut-and-dry you script every time—takes intelligence and meandering, but scoped with tight feedback. That brought us to migrations, version upgrades—Java 7 to 8. Enterprises with massive codebases, 50,000-file codebase, same eight things to change thoughtfully—very repetitive. First success was Nubank, biggest bank in Brazil by market cap at the time—one of these big migrations. Custom Devin extremely optimized for that. As we grew, self-serve and enterprise. From the beginning we'd seen building software in the real world and managing massive products millions use is substantially different from a cute website demo from scratch. In 2026 all Fortune 500 / biggest companies are software companies—Walmart, CVS, JPMorgan, Mercedes. Massive engineering teams. Natural for us—we'd learned early we want real problems people care about.

Today around 75–80%.

A lot of teams on self-serve—grown quite a bit. Startups like Exa, OpenRouter. Ran into someone in my apartment elevator—oh you're the Devin guy, we use Devin.

I've gotten that as well—hey are you Devin? I'm actually not, but that's okay. Why don't we call it Scott? Both self-serve and enterprise are actual engineering teams building real output. We don't focus on individual hobbyists making a cool thing. Real teams building real products they want people to use.

It starts with education. Everyone's gone crazy over AI and agents—buzzwords of the last six months. Still a lot of detail on what deploy actually looks like. We show them, talk through how we partner, direct to right use cases, guidance on maximizing ROI, which projects are or aren't feasible. Enterprises have messy processes. For a massive bank adopting software, repo access, security—typically 12 to 18 months. We work with them to do that as fast as humanly possible.

For Nubank the first one—our entire team was the foreign deployed team. We all flew to Brazil. Agents did not work generally. Sitting with their engineers: this is what you do in that case, what Devin needs to know, read these things, debugging exact problems. We didn't deploy a team—we deployed the whole company.

Let's go through things you think make sense, try some manually ourselves to understand the task, see if we can teach Devin correctly, build the right orchestration—basically building the product almost for one company. It was fun.

Today much more self-start; agents far more capable; figured out more onboarding. Cycles typically 12–18 months—we try to get deployed in like 3 months. Requires them to appreciate priority. If you have 25,000 software engineers in an org costing 10 billion a year that can move three times faster, usually it is a priority. Then security reviews—deploy in private cloud, strict data agreements, tight airwalls.

Typically no. Full foreign deployed motion, but more user education and guidance. We'll fly out, see customers, point to right use cases, teach success, help setup and playbooks. Assist, but your team uses Devin. Incentive-aligned—not hand over a tool and find out. Figure initiatives you care about; projects Devin can make 10x faster now, ones that aren't, workflows instead. More that than using Devin on your behalf.

A couple things really important. One is clearly defining the ROI. Everyone's talking about this last few months—people going crazy on token budgets, one engineer can spend so much—you want to know that's doing real value.

Directionally correct but some places people got carried away. Rank engineers by tokens spent? Rank by how much output and good work is getting done. GPUs are expensive, but if engineers ship three times more, clearly worth it—do it the right way. Tie to specific outcomes—tickets done, projects. Project scoped 18 months for outsourced contractor costing 15 million—do it internally with your team for 1 million in 3 months. Second is being neutral with respect to all the labs, the models themselves. We like being Switzerland.

We're just as incentivized as they are to make their token spend efficient. Devin is purposefully a compound model system—for the right task or even subtask.

For each part of the use case. Tell Devin: customer reported a bug, Jira ticket, solve the whole thing. Investigate the report, reproduce the bug—spin up locally, click through, logs, pinpoint files and command flow, debug, retest, put up for review. Different models good for different parts. Crazy hard tasks want max thinking and the best models. Many tasks are boilerplate—care about fast, immediate, verifiable, cheap and efficient. Not pegged to one model—GPT or whatever only. Use models from Anthropic, OpenAI, Google, our own, open source—dynamically choose for tasks.

In practice there's a lot of positive-sum work together. Software is the only thing we care about. A lot of value in focus and building products and whole motions specifically around that.

Back to why startups ever win. If you do everything, you lose to Microsoft or Google who do everything with trillions more resources, 100,000 more people, infinitely more brand. You build a lasting business by focusing and narrowing on one specific thing, concentrated bet—and your bet has to be right. From there, tight execution. Love Daniel from Spotify's quote. YouTube, Apple Music—why Spotify? I can give other reasons but the truth is we're just going to care way more about music than they are. For us: what does it look like to build software end to end at Goldman or Mercedes—we care so much more. Not a sandbox 30-line algorithm. Messiness of real world, existing codebase, collaborate with humans, plug into ticketing, agent tests own code locally—hairy messy problems we care about. Labs have products; lots of ways to cooperate. Being Switzerland means folks trust us to route to right models, optimize price-performance, direct to right use cases—not incentivized for them to spend more on models; incentivized for value and output.

Even in two years I've heard so many versions of the same argument. Number one pushback: Microsoft already has GitHub Copilot—isn't everybody just going to use that? Reasonable—they owned GitHub, partnership with OpenAI, VS Code. In practice much more innovation to build; Microsoft is a great partner; we do a lot together. 10 years ago: why should Datadog or Snowflake or Databricks exist—clouds care about observability too. In some sense an uncreative way to think. If one end-state future and whoever has most resources wins—we know who has most resources today. But millions of problems, world is dynamic, lots of niches to own, focuses. That will continue. After last couple months of news, Cognition has become a bit more known as the folks betting on independence—high-profile acquisitions happened.

I will not answer that question.

I don't know about dozens but okay.

Folks taking acquisitions—great, reasonable path for exit. For us we came in wanting to build a generational business. Sometimes I see more nihilism: maybe it's too late, not possible anymore—

Maybe not possible to build a new independent business because the labs are going to do everything—all opportunity is taken.

I'd say this even with Devin. If you think going on the Devin web app, giving instructions the way you do now, get the PR, review—that's how software is built forever, then nothing will change. But we have 10 more generations of product experiences to come. Building those is what innovation is.

Not really. People ask before—really would you guys? This is what I'm doing right now. Sometimes I say: we would sell if we thought it was the most ambitious thing to do. Kind of an oxymoron, but genuine. That's what we care about. Funny about money—I don't even have a car.

I rent apartments. I like eating sushi—that's fun. Doesn't cost that much; you can do that off an engineer salary as well.

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

Yeah, I agree. Well, I want to work on Apple. That's what I want to do.

Chicken finger is a great business, but I'm a customer of that business as well.

For us it's very—not rational—but I'd describe it: one thing if we tried, gave it our all, and just weren't good enough. I'd be salty as hell—wouldn't be fine—but an outcome I could live with. If we felt we could have gone for it all, pushed harder, and didn't—I don't think we'd live with ourselves. Why so excited to do this, spend all this—we want to achieve our potential and build what we were meant to build. Maybe something, maybe nothing—but you'd rather find out and see.

Yeah. Thanks for having me.