Scott Wu

They called third place, then second, then first, and none of the names was his. He was seven or eight years old, entered in a seventh grade math competition, a middle school contest held at the local university, and what he kept from the afternoon was not the result but the fury: "I just remember being so pissed about that."¹

Wu is the co-founder and chief executive of Cognition, the company behind Devin, and the word he uses for that afternoon is his own. Salty. Asked what it means, he gives a definition with no self-flattery anywhere in it, "salty just means that you take offense to the idea of losing," and then refuses to account for it: "I can't really give you a rational explanation for why it is."¹ Two things are true of him at once, and the pair organizes everything below. The drive is total on his own accounting. Asked how much of his brain runs on competition he answers "it's all I do," and describes building a company in the vocabulary of a game he has been playing since he was small: "a tree search where you're exploring the different options in the decision tree and you're trying to figure out how to lead to victory. That's like the only thing I do in my life."¹ The loss is not total. He carries a stated ceiling on what a defeat is allowed to cost him, and the record reads differently once that ceiling is in view. The standing caveat is the shape of the evidence: one long interview, conducted by someone who says he met the founders in 2023 when the company had no revenue and has talked with Wu repeatedly since, in which nearly every number is Wu's own.

The ladder

The competitions were not a hobby alongside a childhood. They were the childhood. "My whole goal was to become world champion of competitive programming," he says of the years before he could drive.¹ What he praises about that world is a piece of structural design rather than a feeling. You compete at your school; do well enough and you qualify for the city, then the region, then the state, then the national contest, then the international one. The ladder's real function, in his telling, is sorting: "sooner or later you get to kind of meet people who are like you."¹ The people he met in those national and international rounds were, he says, more his childhood friends than the people around him in Baton Rouge, Louisiana, were.¹ Hold onto that, because it is the first place he separates from the archive's other pure competitors. The contests did not cost him a social life. They issued him one.

He collected formal games the way other children collect cards: Super Smash Bros. Melee, with tournaments; Tetris; a lot of poker; some chess, about which he is unsentimental, "I was okay at chess, I was not good"; and some Go.¹ Go is the game with a family history attached to it. His father was a seven dan player, which Wu converts for the listener at roughly a 2,300 or 2,400 chess rating, and who played a professor at his graduate school in China on weekends, generally won, and talked the games over afterward. The professor moved to the United States to teach, wrote back urging Wu's father to follow, and offered to help with the visa and the applications. His father applied to graduate school in the United States; Wu was born after the move. "My parents came to the US in some sense because of Go."¹

The wall

The other half of the household was his mother, and this part of the record has to be handled carefully, because the evidence and the reading laid over it are not the same object.

Wu's own testimony is narrow and concrete. She was the most salty person in the family: "she would always be, oh no, no, I'm better at this, I can beat you at this."¹ She played on her school ping pong team and studied some math, but the competitiveness was "more her personality than any one thing that she really put all of her competitive energy into."¹ Asked directly whether she told him he was the best before there was evidence for it, he agreed: "even when I was tiny, she would tell me that I was extremely talented."¹ And the household made the claim structural. There were no pictures of his parents on the walls. There were old math competition trophies, hung as each one arrived, and his account of the intent is that this was deliberate, that the trophies were the thing the household valued and that a trophy on the wall existed to be replaced by a better one.¹

The formation read off that evidence belongs to the interviewer, not to Wu. David Senra, who has spent roughly a decade reading founder biographies, offers a two-part taxonomy of the parents in them: the contempt case, a father telling a son he will never amount to anything, with Larry Ellison and Elon Musk as the standing examples, and the belief case, with Estee Lauder as his. He places Wu's mother in the second and invites agreement.¹ Wu does not supply it. What he offers instead is a qualification: "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."¹ He confirms the belief and declines the causation, which is a distinction worth more than most of what founders say about their childhoods. The reader is directed to the file on belief before evidence, which sets the pattern against the contempt case and closes on the caveat that governs both: each type is reconstructed entirely from the children it happened to work on.

The governor

The interview puts a specific proposition to him. Larry Ellison has said he is addicted to winning but fears losing more than he loves winning; asked about it, Michael Dell agreed that the pain of losing exceeds the pleasure of winning.¹ Wu accepts the description of the feeling and then immediately fits it with a governor: "to get anywhere, you got to lose a lot," and "losing feels way worse than winning feels good, but not by enough that it makes me want to stop trying."¹

That second clause is the least quotable sentence in the conversation and the most load-bearing. Aversion to losing is not rare, and by itself it predicts nothing, because its far more common output is a person who quietly stops entering. The trait only becomes an engine inside a narrow band: large enough to produce real effort, small enough not to produce avoidance. Wu is describing where the needle has to sit. The file on fear of losing as fuel holds the trait itself; the file on playing to win holds Marc Rowan using nearly the same words to diagnose institutional decay, which is what the identical emotion does to an incumbent with something to protect.

Thirteen percent, and the eighty seven

Cognition got started in early 2024, and Wu's own summary of the position is not flattering: "dude, we were late by a lot."¹ OpenAI had started at the end of 2015, Google's research labs were more than a decade old by his own reckoning, and in code specifically, GitHub Copilot had been in engineers' hands for years in exactly the autocomplete and question-and-answer form Cognition was betting against. His generalization from that seat is blunt: "when you start a company, you kind of have nothing. You have no right to exist," no resources, no people, no brand awareness, and therefore "no right to win." What sometimes converts that into a win is a single move: "plant your flag in the ground and put a stake into what you think the future is, and you run like hell towards that."¹

The flag went up in March 2024. It went viral, and it arrived with no customers and no revenue behind it.¹ The number attached to it was a score on SWE-bench: 13 percent, against a best known result he puts at 3 or 4 percent, which is also a failure rate of 87 percent.¹ The reception came in at both poles, the coolest thing ever on one side and "there's no way this is ever going to work, this is totally a scam" on the other.¹ The reason he shipped anyway was not the level but the derivative, and he can date the moment he started reading it that way: the first real task Devin completed was setting up MongoDB, working through the error, search, retry loop an engineer would otherwise run by hand, and "I could not sleep that night."¹ The reader is directed to the file on evals and benchmarks, where this is the clean case of a score used rhetorically and correctly, since a benchmark whose state of the art sits at 4 percent is measuring something almost nobody can do.

He concedes more of the criticism than the founder-under-fire genre usually allows. "Definitely we were early, which is a very fair criticism," and "we were wrong on a lot of things," and, on the hate, an admission that he may have under-updated: perhaps they should have done something in the middle, "maybe more than we should have. I don't know."¹ One thing he does not reconcile. "We were late by a lot" and "we were the first ones that actually planted that flag" are both offered as explanations of the same outcome, and the reading that would join them, late to AI and first to agents as coworkers, is never stated on tape.

The whole company flew to Brazil

The launch produced attention and no business. People asked for the product, Cognition scrambled together pilots, and, in his words, "perhaps unsurprisingly, they were all just failing."¹ By April and May of 2024, running agents on the models of that moment, the team stopped asking how to make Devin good and asked instead which task would be the first one worth handing over at all. The answer they reached was neither the hardest work nor the easiest: not deep architecture, which needs all the context, and not anything a script already handles, but work "repetitive enough and scoped enough and on a tight enough feedback loop" that an agent could diagnose and fix it end to end.¹ In practice that meant migrations and version upgrades, a Java 7 to Java 8 move of the kind enterprises run across a 50,000 file codebase where the same eight things change in every file.¹

The first success was Nubank, then the largest bank in Brazil by market capitalization, on one of those migrations, with a Devin custom built and, he says, extremely optimized for the job. Asked whether he sent a deployed team, the answer is the operational form of everything above: "we didn't deploy a team. We deployed the whole company." The entire team flew to Brazil and sat with Nubank's engineers, working out case by case what Devin needed to know and read, and building the product for one customer.¹

That work set the shape of the business. Enterprise is 75 to 80 percent of revenue, with Goldman Sachs, Mercedes, and parts of the United States government among the customers and Exa and OpenRouter among the self-serve teams; the vendor adoption cycle a large bank normally runs at 12 to 18 months, Cognition tries to compress to about three.¹ The interviewer puts the climb at roughly a million dollars of revenue to roughly five hundred million in about twenty months, hedged as an approximation, and Wu does not dispute it.¹

Switzerland

Two structural choices follow, and both are arguments about incentives rather than about technology. The first is model neutrality, which he calls being Switzerland. Devin is built as a compound model system that routes each subtask to a different model, drawing on Anthropic, OpenAI, Google, Cognition's own models and open source ones, choosing maximum reasoning for the genuinely hard steps and the cheapest fast answer for boilerplate that can be verified.¹ The claim underneath is about alignment: "we're not incentivized for them to spend more on the models, we're incentivized for them to get value and to get output out of it."¹

The second is focus, and it is the reason he gives for why any startup ever beats an incumbent: "if you do everything, you will lose to Microsoft or Google," who also do everything but with trillions of dollars more and a hundred thousand more people. The alternative is to narrow to one thing and make a concentrated bet on it. His load-bearing citation is Daniel Ek on why Spotify survived Apple Music, which he quotes as: 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.¹ Wu's version of caring more is unglamorous and specific: what it actually takes to build software end to end at Goldman Sachs or Mercedes, plugging into the ticketing system, giving the agent the ability to run and test its own code, working with a codebase as it exists rather than a sandbox problem with a thirty line answer. The reader is directed to the file on singular product focus, where the same constraint is held by an operator with no technology in it at all.

The same incentive logic drives his one public disagreement with the way the industry measures itself. Ranking engineers by token spend measures consumption, an input, and quietly rewards more of it: "let's try and rank people by how much output they're actually producing, or how much good work is actually getting done." The measure he offers instead is a project scoped at 18 months and 15 million dollars with an outsourced contractor that his customer's own team finished internally in three months for one million.¹ The two files on that argument, output over token spend and token spend management, take opposite sides of it, and the second one notes what Wu's own position is worth given that a company absorbing tokens into its product margin has an interest in the number receding from view.

The mission

The turn is where the competitor's frame runs out, and he gets there by refusing to pattern match. His account of forecasting is that people extrapolate from what they have seen, which works "99 percent of the time," and that the exceptions are the periods when something real has changed, where the only usable method is first principles.¹ The measurement he reasons from is METR's: a couple of years ago an AI could sustain roughly 10 to 20 seconds of unassisted human-equivalent work before someone had to correct it, and that duration has been doubling roughly every couple of months, reaching hours.¹ The blunt question follows: why not days, why not months, and what does the world look like when everybody has an agent that can do months of work at a stretch.

From that he reads off a ladder of delegation, which is his real contribution and is held in the files on the agent time horizon and on commissioning versus steering. Seconds of unassisted work means you are handing over a command. Hours means you are handing over a task. Years means you are handing over a mission, and the only thing left in human hands is what the human cares about: a societal problem worth coordinating on, a game worth designing, a line of materials science worth pursuing on the agent's own timeline.¹ The reference point for what such a mission looks like comes from the interviewer, not from Wu, and it is Edwin Land setting a researcher on one question, how to bring color to instant photography, which the man reportedly spent two years thinking through before solving it.¹ Wu takes the frame without hesitating.

Underneath sits the claim he would most like to be right about. Software only gets built today if it will be used enough times to repay a team of engineers, "at least a million times or something, and maybe it's 10,000," which is why an enormous amount of white collar work stays manual: reading fifteen profiles, filling in forms, assembling one spreadsheet from one afternoon's research.¹ Nobody staffs a team for software that gets used once. If the threshold falls to one, that entire category converts, and programming stops being a craft and becomes what he says it always was, the interface: "Devin is the way that humans can tell their computers what to do."¹

Notice what that does to the man who was formed inside scored, finite contests where the placings are read out the same afternoon. The end state he is building toward has no placings in it. A mission runs for a year and returns something nobody ranks. He never remarks on it, and the record does not close the gap, but it is consistent rather than contradictory, because his own success condition had already stopped being a placing.

What he will not trade

Cognition has become, he says, somewhat known as the company betting on independence, which is a position defined against a season of high-profile acquisitions.¹ He will not say how many offers have come in; asked, he answers flatly, "I will not answer that question."¹ What he will state is the condition: "we would sell if we thought it was the most ambitious thing to do," which he immediately concedes sounds like an oxymoron.¹ It is worth being exact about what that is and is not. It is not a vow of ownership. It routes the decision through ambition rather than control, which leaves a sufficiently ambitious buyer as a live case in a way it is not for Todd Graves, the only entrepreneur from Baton Rouge wealthier than Wu, who has turned down acquisition offers on the flat ground that it is not about the money.¹

The rest of the evidence for the claim is domestic and cheap to verify. He does not own a car. He rents an apartment. His stated indulgence is sushi, which he notes is affordable on an engineer's salary.¹

Then the close, which is the ceiling stated plainly. If they tried, gave it everything, and simply were not good enough, that "would be fine," a formulation he corrects on the spot, because he would be "salty as hell" and it would not feel fine at all, but "it would be an outcome that I could live with." The unlivable outcome is the other one, the loss where they could have pushed harder and did not: "I don't think we would live with ourselves in that outcome."¹ What he wants, in his own summary, is to "achieve our potential and build what we were meant to build," and "maybe that's something, maybe that's nothing, but you'd rather find out and see."¹

Under examination

Wu carries no plate here, and the absence is a finding rather than a gap.

Half of the case for one is as strong as any in the catalog. The drive is articulated by the man himself, in his own vocabulary, with a dated first instance, a definition, and a calibration, which is exactly the standard a plate is held to. The archive already holds it, under fear of losing as fuel. The other half does not reach the same bar. That his mother declared him the best before there was evidence is a fact he confirms; that the declaration is what formed the founder is a reading he was offered and specifically declined, and the archive does not promote an interviewer's synthesis into a subject's self-knowledge.

The near miss is worth naming because it is instructive. Set beside the Competitive Ascetic, Wu's drive is the same mechanism stated in almost the same words, Phil Knight's "I simply didn't want to lose" against Wu's "you take offense to the idea of losing." Everything else diverges. That plate requires a founder who discards social life as surplus; Wu's competitions were where he found his friends, and the company he built is the same room enlarged. That plate requires losing to register as something close to death; Wu has a ceiling, says so unprompted, and locates the unbearable outcome in effort withheld rather than in defeat. Two founders can share an engine and build nothing alike, and the difference here is a cause upstream of the drive rather than a mechanism downstream of it, which is not what this taxonomy classifies. The honest record is a founder filed under examination, two patterns filed where they can be argued with, and no third thing invented to hold them together.

Key facts

• Co-founder and chief executive of Cognition, the company behind Devin, which got started in early 2024 with a founding team of nine, most of whom had founded companies before.
• Defines the trait he is known for himself: "salty just means that you take offense to the idea of losing," first evidenced at seven or eight by placing nowhere in a middle school math competition.
• Attaches a governor to it that most competitors do not state: losing feels worse than winning feels good, "but not by enough that it makes me want to stop trying."
• Grew up in Baton Rouge, Louisiana, aiming to become world champion of competitive programming, and collected Melee, Tetris, poker, chess and Go alongside math and programming contests.
• His father was a seven dan Go player whose weekend games with a professor in China produced the invitation that brought the family to the United States.
• Confirms that his mother told him he was the best before there was evidence, and that math competition trophies hung where family photographs would normally go, while declining the claim that his parents steered him toward founding anything.
• Launched Devin in March 2024 with a SWE-bench score of 13 percent against a prior best of 3 or 4 percent, with no customers and no revenue, and took criticism from both poles.
• Reached the first working use case by narrowing to repetitive but unscriptable work, migrations and version upgrades, and won Nubank by flying the entire company to Brazil rather than a deployed team.
• Runs Devin as a model-neutral compound system, routing subtasks across Anthropic, OpenAI, Google, Cognition's own and open source models, and argues engineers should be ranked by output rather than by token spend.
• Reads METR's measurement of unassisted agent work, roughly 10 to 20 seconds a couple of years ago and doubling every couple of months, into a ladder of delegation: seconds is a command, hours is a task, years is a mission.
• Owns no car, rents an apartment, and states the sell condition as a conditional rather than a vow: "we would sell if we thought it was the most ambitious thing to do."