Devin Will Not Replace Engineers. It Will Multiply Them, Says Cognition’s Scott Wu

Sarah Guo and Elad Gil introduce Scott Wu as both a very good human software engineer and the co-founder and CEO of Cognition, an AI lab that calls itself focused on reasoning. The company had just released a demo of Devin, an AI software engineer that can increasingly handle entire tasks end to end, with leading results on SWE-bench, a software-engineering benchmark of real-world GitHub issues. Guo says the demo “broke the internet, at least among tech Twitter.”

The pitch in the intro is not a better autocomplete. Devin is framed as something that can take a job from prompt to finish — and the conversation is about what that does to the people who currently hold those jobs. Wu is also, the hosts note, a world-class competitive programmer who grew up in the International Olympiad in Informatics circuit. That biography is not color. It is the through-line of how he thinks about hard problems.

Wu always loved math. His older brother Neal was the first to get him into programming; he learned around age nine and fell in love with taking ideas and making them real. Math and programming contests became his entire life through school — what he thought about in the shower, how he framed every problem he ran into. By the end of high school he was set on going into tech.

IOI, he explains, is the Olympiad of code: every country sends its best, there are gold, silver, and bronze medals, and the work is algorithmic problem-solving plus implementation. Practice is not a secret. It is a lot of effort, analytic thinking, and being “brutal about your own shortcomings.” There are standard algorithms — shortest path, binary search trees — but every contest problem is unique. The beauty is creative problem-solving: reducing a new task to shortest path, or modifying an old algorithm for a new case.

He grew up in Baton Rouge, Louisiana, where almost nobody shared those interests. Competitions, held once a year plus the odd training camp, were the first time he met people who did. For most of the year the community lived online, talking problems and becoming close friends. Guo notes that this is how she met him — through a high-school classmate she had backed — and that he knew Scale AI’s Alexandr Wang in middle school. A lot of that tight-knit group became founders and went into AI.

Wu argues that competitive programming has more in common with entrepreneurship than it looks. You are not literally using shortest path to build a company. You are hunting the assumption everyone takes for granted that might not be true, and finding the unintuitive idea hiding behind it. “A lot of building a great startup is about finding something that the world doesn’t believe is true yet but actually is,” he says, plus first-principles thinking and the habit of pushing yourself to improve. Guo ties that directly to Devin: people saw behavior they did not think could exist yet, and many were waiting for the next foundation-model leap instead.

Cognition got started around November. The initial group was largely Wu’s friends from the math and programming contest community, each of whom had since taken a different path through AI. Wu had run a machine-learning company, Lunchclub, for about five years. Co-founder Walden Yan had built a lot of early Cursor. Co-founder Steven Hao was one of the first engineers at Scale AI. The whole founding team had grown up on problem-solving and then spent five or ten years inside AI and AI-infra companies.

Why code? Two reasons. It has a special place in their hearts. And Wu would even say that most of the progress in the world in the last 20 or 30 years has come from software. “It’s crazy when you think about how actually everyone everywhere needs more software engineering, not less.” Every company is hiring more engineers; everyone has more ideas than time to build them. Accelerating the pace of code, they felt, was a way to accelerate everything else.

Wu’s definition is direct. Devin is an AI software engineer, meaning it is fully able to make all of its own decisions the way a human engineer would. Writing and editing code is a huge piece of that, but so is working with the command line, using the browser, reading documentation, deploying, testing, and debugging. You give it a simple prompt of what you would like to build; it goes through the whole job end to end. You can also look at what it is looking at, give feedback, and steer. The intended feeling is looking over another engineer’s shoulder.

The online reaction was bigger than the team expected. Wu even found his middle-school math videos circulating on Twitter again. The comments split the usual way: skeptics who said there was no way this could work, and people who said everyone was about to lose their jobs. Wu’s answer to the second camp is the opposite. “If anything there’s going to be more engineers, not less.”

First, demand is underestimated. A lot of Excel and cobbled-together tools exist because engineering is hard. There are so many problems that could be solved with code that multiplying every developer, he thinks, gives you more developers, not fewer. Second, Devin is “very much not the type to decide what to do.” The average software engineer today might spend 20 percent of the time on fundamental problem-solving and 80 percent writing it in code. With Devin they should be able to do 5x more and spend all of their time on the creative part.

Gil offers a loose analogy: human calculators. You probably do not want people doing that work, and taking it away lets more capable people expand into something else. His core optimism about AI is that it is democratizing — tapping latent demand that was unaffordable, where software did not match the workflow, was not good enough, or was not sophisticated enough. Wu agrees: that is largely the case in knowledge work everywhere. Enabling every human to do more only helps the whole stack accelerate.

Guo has started seeing Devin-like UIs everywhere. Earlier agents would go off and come back 30 minutes later with something you did not like — “really eager interns” with the right intent and the wrong path. Devin’s four tabs, she notes, put planning, the shell, the code, and the browser on the table so you can steer. She calls it a powerful paradigm that is catching on. Wu says the team built for themselves: they are all software engineers, they use Devin while building Devin, and they worked backwards from how they would want to interact with an AI coding teammate.

The intern analogy is the design constraint. If an agent takes the task and disappears, it has one try to do everything perfectly. If you have a human intern on a project for a day, you check in four or five times, spend 30 seconds on feedback, and they both learn and still deliver valuable work — much faster than doing those pieces yourself. Transparency is not a nice-to-have skin. It is how you get more than one shot at being right.

Asked what Devin is already better at than humans, Wu starts with encyclopedic knowledge, then DevOps and environment setup, which he thinks is honestly hard for people. One of the first exciting internal moments: the team was stuck for about an hour getting a database spinning and Kubernetes up, lost in errors and stack traces, and they asked Devin to set it up and tell them how because they could not. It did. The step-by-step loop of running commands, reading errors, opening a port, installing a package, maps cleanly onto an agent workflow. Data analysis is another: find a dataset, pull a CSV, sanitize, run the analysis, build a visualization, start to finish.

The weakness is the complement of the job thesis. Devin is not going to decide what to do. You provide the precise formulation of what you want built; Devin does the thoughtful execution. You would not go to it and say “build me a great business.” Wu will not go deep on the internals, but he offers a frame. A perfect intelligence, given a GitHub issue, could stare at the codebase and emit a clean diff. Practically, humans and AIs today take the messier path: run the code, reproduce the bug, add print statements, read logs, ask Stack Overflow. A lot of what Cognition works on is getting Devin to plan and evaluate that way.

On a five-to-ten-year horizon, Wu thinks what we call software changes a lot. He sees Devin as a step toward the next generation of human-computer interface. Ten years from now, people will look back and find it crazy that you had to learn esoteric languages and work through stack traces just to communicate with a computer. Software engineering is already a discipline about making the computer do what you want. The fun part — taking a problem, figuring out the form of the solution, the edge cases, the flows, the architecture — stays with people. The Devins of the world take that to implementation.

The demand number he keeps coming back to: about 30 million software engineers in the world, which is still only 0.4 percent of people. Give those people the ability to do much more, and enable so many others to work with software, and the market does not look small. In the next one to two years, engineers who are adaptive on this trend will multiply themselves. Hardware, foundation models, and the agent work Cognition does are all rising tides. Asked what will improve agents most — better base models, reasoning, memory, self-play, tool use — his honest answer is yes to all of the above, including inference speedups. A perfect intelligence would not need tools; a perfect tool-user might not need that intelligence. In practice every layer is a race.

There is a meme that the hottest new programming language is English. Wu thinks there is a lot of truth to it, and that the fundamentals stay valuable anyway. Everyone uses the internet; people who work with networks still benefit from understanding TCP. Logic gates, algorithms, how computers work, logical reasoning — those stay useful. The role five or ten years from now looks like a mix of today’s technical architect and product manager: take the business problem, break down what the solution should be, and let the agent write the code.

Gil asks the longer question: five years ago he would have told kids to study computer science and math; 20 years out he is less sure. Wu says the experience is worthwhile even if it is not practically useful, and that the fundamentals will stay useful for a long time. Superintelligence and the singularity are hard to predict with high confidence; everyone in AI has made their guesses. What he is willing to say is that AI’s concrete impacts on work, the economy, and people’s lives will show up a lot sooner than that. ChatGPT was about a year and a half ago. There is still a lot to do on the research side and the product side.

Guo had lunch that day with one of Cognition’s new hires, who told her it just made sense that “Scott Wu is going to be the last human software engineer slash labeler.” The joke is that a world-class engineer’s process becomes training data. Wu laughs it off the only way he will: good reasoning, good fundamentals, and a good understanding of technology are going to be useful for quite some time. Maybe eventually you upload knowledge into your brain and become a perfect rational decision-maker. With what exists today, the logical fundamentals stay relevant.

On hiring, he is proud of the early team. The last five or six months have been about the size of the problem: software is a huge driver of change and a deep, interesting problem. Almost everyone at Cognition was either a founder before or was explicitly thinking of starting a company if not for this one. He wants high ownership, people focused on outcomes rather than politics, plus creativity, work ethic, and communication. So far they have grown mostly through a close network of people they already knew and worked well with.

Who are they looking for now? Great engineers and researchers. The company is still very early and the mission is big, so the archetype is a strong technical thinker who is excited to work on multiple parts of the business — customers, strategy, research, engineering, product. In an early startup, he says, they are kind of all doing everything. That is the same contest-kid habit in a new form: pick a hard problem, refuse the office-politics version of the job, and stay close to the result.