Part IThe Story
Here's Who Got Third Place
"Here's who got third place."
The room was at a local university in Louisiana, borrowed for an afternoon by a middle-school mathematics competition, and the names were being read aloud in ascending order of glory, the way such names always are. The contestants were seventh graders, mostly, eleven and twelve years old, sent by their schools with pencils and calculators. One of them was not a seventh grader. He was in the second grade, seven or eight years old, and he had come because his older brother did these competitions and because, as a younger brother, he went where his older brother went and tried to do what his older brother did. He had taken the test. He was now waiting, with the particular certainty of a small child who has done a thing well, to hear his name.
"Here's who got second."
He was not third. He was not second. He was not first. The ceremony ended and the children dispersed, and two decades later, after he had become the most decorated American in the history of the International Olympiad in Informatics, after he had built and left one venture-backed company and founded another that would be valued at $26 billion before its third birthday, Scott Wu could still summon the feeling with no apparent effort. "I just remember being so pissed about that," he told the podcaster David Senra in June 2026. He did not offer an explanation. He did not seem to think one was available. "I can't really give you a rational explanation for why it is."
The next year he entered the same competition in the ninth-grade division. He was in the third grade. He got a perfect score.
That is one way to begin a story about Scott Wu, and it is the way he himself tends to begin it, which should make a writer slightly suspicious. Origin myths told by their subjects have a way of being too clean. But this one has the unusual virtue of not flattering its teller. The salient fact is not that he won the following year. It is that a child lost a contest he had no business entering, against opponents half again his age, and responded not with relief at having been allowed to compete but with fury at having failed to win. Wu calls this quality being "salty." He traces it to that afternoon. He says he has no memories from before it.
The other way to begin is with a different date.
On November 17, 2023, the board of OpenAI fired Sam Altman, and for roughly five days the entire technology industry did nothing but watch. It was the most consequential corporate drama of the decade, a fight over who would control the most valuable research lab on earth, conducted in public through cryptic posts on X and leaked accounts of weekend board calls. On that same day, according to an account in Colossus magazine, a 26-year-old who had buried his mother a few weeks earlier—lung cancer—founded an artificial intelligence company. He named it Cognition. Its stated purpose was to teach a machine to do the one thing on earth that he, by any reasonable measure, did better than almost anyone alive.
The boy who could not abide hearing other children's names called would spend the next three years building a thing whose name, Devin, would be called in boardrooms at Goldman Sachs and Mercedes-Benz and the United States Army, on a sign in Times Square, in the comment sections of his own LinkedIn posts, where the company's official account would occasionally chime in like a precocious child: Hey that's me!
By the Numbers
The Cognition Ascent
3IOI gold medals for Wu (2012, 2013, 2014), first overall in 2014
10IOI gold medals held by Cognition's founding team
$350M → $26BValuation from early 2024 seed to May 2026 round
$492MRun-rate revenue announced May 2026, from $1M in September 2024
89%Share of code committed internally at Cognition shipped by Devin
72 hoursTime from Google's Windsurf deal to Cognition's handshake agreement
The Little Brother's Problem
Baton Rouge is not a city that produces technologists so much as one that refines hydrocarbons. The capital of Louisiana sits on the east bank of the Mississippi along the industrial corridor that runs down toward New Orleans, a hundred-odd miles of refineries and chemical plants whose flare stacks give the river its night light. Wu's parents had come from China for graduate school in chemical engineering and had stayed, by one account settling into the oil-and-gas economy in the work of air-emissions permitting—the paperwork that stands between a plant and the atmosphere. It is hard to imagine a profession more remote from the romance of Silicon Valley, or one more representative of a particular kind of immigrant bargain: technical competence, deployed quietly, in service of an industry that does not know your name.
What the household valued, in the accounts that survive, was intellectual work and the evidence of it. Wu told Senra that in the family home the math trophies came to take the place of family photographs. He offered this as a fact rather than a complaint.
Neal was the first son. He was four or five years older, and he started doing math competitions in middle school, in the sixth grade, at around the time Scott was in the first. "Naturally I, as a little brother, would go and just watch what he was doing and try to learn some of the same math too," Wu has said. They played video games together, and the younger brother lost, and hated losing, and the pattern of the next decade was set in the living room: an older sibling who was very good, a younger one who could not stand being second.
Neal would go on to win three gold medals at the IOI himself, a team silver at the International Collegiate Programming Contest, and second place at the 2012 Google Code Jam; he would graduate from Harvard in 2014, work at Facebook and Google, and serve as a head teaching fellow at Harvard before joining his brother's company. He was the trailblazer, in other words, and he was also the one who introduced Scott to programming contests, by way of the USA Computing Olympiad. The record suggests a sibling relationship of unusual generosity on one side and unusual hunger on the other, and that both turned out to be useful.
Scott learned to program at nine, writing small programs in BASIC on a TI-83 graphing calculator—the beige-and-gray brick that generations of American students have used to cheat on algebra tests by storing formulas in memory. He used it instead to make things. "The thing I loved about programming at nine," he wrote in 2026, "is the exact same thing I love about it now. You have an idea, and you get to see that idea become reality."
He was, by then, a known quantity on the Louisiana competition circuit. In 2011, as an eighth grader at Glasgow Middle School in Baton Rouge, he won the national championship of MATHCOUNTS, the middle-school competition whose televised final round—a buzzer-driven "Countdown" of rapid-fire questions—is the closest thing American mathematics has to Jeopardy! A clip of Wu from the 2010 competition would resurface in March 2024, after Devin's launch, and collect millions of views. In it he answers questions before the moderator has finished reading them. The effect is less of calculation than of recognition, as if the problems had been written in a language he simply spoke.
What is easy to miss in the clip is how alone he looks. There weren't, in Baton Rouge, many children like him. He has said so directly.
There weren't many other kids around me who loved math and programming, so I found my people in programming competitions, traveling around the state, then the country, then the world to sit in rooms with other kids obsessed with the same hard problems.
That sentence carries more weight than its plainness suggests. A child of immigrants in a petrochemical capital, unusual among his peers, discovers that his peers exist—they are just scattered across the state, then the country, then the globe—and that the only place they convene is a room with a problem set on the table and a clock on the wall. Belonging, for Wu, was not a neighborhood or a church or an ethnic enclave. It was a contest. You found your people by beating them, or trying to.
— Scott Wu, to David Senra, June 2026It's like 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 the only thing I do in my life.
A Country of Rooms
The International Olympiad in Informatics is, for the uninitiated, a week-long competition for high-school students in which each participating nation sends a team of four, and each contestant sits two five-hour sessions solving algorithmic problems of a difficulty that most professional software engineers will never encounter in their working lives. Roughly the top twelfth of contestants receive gold medals. To win one is to be, in a meaningful sense, among the best young programmers on the planet in a given year. To win three is rarer. To finish first overall is the kind of thing that happens to a nation perhaps once a generation.
Wu won gold in 2012, in 2013, and in 2014. In 2014 he finished first in the world.
That same year, in the same competition, a student named Steven Hao also took gold, placing sixth globally. Hao would go on to study computer science and mathematics at MIT from 2014 to 2018, intern at D.E. Shaw and Dropbox, and join Scale AI in 2018, where he became one of the company's top engineers—a trajectory of quiet excellence inside other people's institutions. Nearly a decade after they shared a medal podium, he would become Cognition's chief technology officer. The company was, in one sense, already assembled at the IOI in 2014. It just didn't know it yet.
This is the part of the story that the venture press tends to compress into a list of accolades, and the compression loses something. The rooms where Wu found his people were not merely venues for prizes. They were, it now appears, the incubator of an entire industry's leadership. Most of the people running the largest AI companies met as teenagers, Colossus observed, competing for their countries on international math and science teams—and it named, among others, OpenAI's Greg Brockman, Anthropic's Dario Amodei, and Alexandr Wang, who built Scale AI and went on to Meta. Wu, writing in 2026, made the same observation from the inside: "A lot of those friends now run or do research at the top AI companies in the world."
It is a curious historical pattern. The adolescents who spent their summers in hotel ballrooms in Taipei or Almaty or Bali racing to implement segment trees became the adults who decide how the most powerful technology of the century gets built. They know each other. They have competed against each other. Some of them have very long memories about who beat whom. Colossus reported that among this cohort, most agree the von Neumann was Scott Wu—a phrase that should be read with the knowledge that Colossus counts Cognition as a sponsor, and also with the knowledge that people who have lost to Wu at things tend to say similar things unprompted. Isaac Kim, who joined Cognition in 2025, has described watching him solve a Putnam problem aloud in ninety seconds without touching pen or paper.
Wu's own account of what he loved about the competitions is revealing for what it emphasizes. "I find programming contests to be a really fun mix of innovation (coming up with new algorithms or discrete math) and engineering (coding/debugging skill, practical understanding of systems/compilers, code optimization)," he once wrote. "I also really enjoy the strategy component of programming contests, which is very unique compared to e.g. math or physics olympiads." Strategy: which problem to attempt first, when to abandon a dead end, how to allocate five hours across three problems whose difficulty you can only estimate. A programming contest is a resource-allocation problem wrapped around a set of algorithmic problems. The best competitors are not merely the smartest; they are the ones who triage best under a clock.
He kept competing long after it could plausibly help his career. He took third at the Topcoder Open Algorithm final in 2017. He took third at the Google Code Jam in 2021, by which point he was chief technology officer of a venture-backed startup with tens of millions of dollars in funding and, presumably, other things to do. His peak rating on Codeforces, the Russian-run platform that serves as the sport's global ladder, reached 3350—"Legendary Grandmaster," the platform's highest tier, rendered in a red so deep that aficionados describe the top of the scale as approaching the color of Nutella. His handle,
scott_wu, is unadorned. He did not need a pseudonym. Everyone knew.The Matchmaker
The years between the last Olympiad and the first company have the texture of a prodigy trying on adulthood and finding most of it too loose.
In May 2014, the spring of his IOI win, Wu went to work as a software engineer at Addepar, the wealth-management software company co-founded by Joe Lonsdale, and stayed until August 2015. Lonsdale—Palantir co-founder, serial company-builder, later the founder of the venture firm 8VC—would later describe him on his podcast as having been "a top young talent at Addepar," and would lead Cognition's funding round at a $4 billion valuation in March 2025, a decade after the teenager left his payroll. Another account has Wu as one of the first interns at Hudson River Trading, the high-frequency firm that has long recruited from the Olympiad pipeline. Both are the natural destinations of a teenager like him: places that pay very well for the ability to think quickly about systems.
Then Harvard. Here the record forks, and I would rather show the fork than pretend it isn't there. Wikipedia has Wu attending for two years and dropping out. Other accounts have him graduating with the class of 2019 with a degree in economics. Everyone agrees on the economics—not computer science, a choice that tells you either that he felt he had nothing left to learn about computer science in a classroom or that he wanted to understand the world he would eventually sell to. Everyone agrees that as a sophomore he represented Harvard at the 2016 ICPC World Finals, with Calvin Deng and Johnny Ho, and that the team took gold, finished as North American champions, and placed third in the world. A photograph of Wu from those finals is the one Wikipedia uses. He looks about fifteen.
In June 2017, he co-founded Lunchclub.
Lunchclub was, on its surface, an unlikely vehicle for a competitive programmer. It was a professional-networking service that used machine learning to match people for one-on-one meetings—coffee, initially, and later video calls—on the theory that the right introduction could change a career. Wu's co-founder was Vladimir Novakovski, himself a former member of the 2000 U.S. Physics Olympiad training team, which is to say a fellow alumnus of the rooms. The company raised a $4 million seed and a $24 million Series A, drew backing from Lightspeed, Coatue, and Andreessen Horowitz, and put Wu on the 2020 Forbes 30 Under 30 list for consumer technology.
It is worth pausing on what Lunchclub actually was, because the irony is too good to step over. A young man who had found his people only in contest halls, who had grown up the unusual child in a city with few others like him, spent five years building an algorithm whose purpose was to find people their people. The matching engine was his. Lunchclub's premise was that belonging could be computed—that the right two strangers, correctly paired, would recognize each other across a table.
Whether it worked is a matter of some dispute and does not much matter here. Wu was its CTO until June 2022, when he left, by his own description, to pursue different interests. The phrase is the standard departure language of the startup world and conceals more than it reveals. Lunchclub's model, built on in-person serendipity, had been reshaped by a pandemic that made strangers unwilling to meet. Wu, for his part, had been competing in the Code Jam the year before. One gets the sense of a builder whose company had become a job, and whose real attention was elsewhere.
Asked later how a person knows when to move on from an idea—a question South Park Commons, where Wu spent time after Lunchclub, put to him directly on its Minus One podcast—he spoke about it in terms that would apply equally well to abandoning a problem in the third hour of a five-hour contest. You estimate the expected value of continuing. You compare it to the expected value of what else you could do. You are brutally honest about which number is larger. "The willingness to be very brutally honest with yourself," he has said of what competition taught him, "the ability to just examine your own flaws."
November
There is a sentence in Isaac Kim's account of working for Wu that stops the reader cold. It arrives in the middle of a LinkedIn post that is otherwise the standard corporate encomium—smartest person I know, we're hiring—and it reads: "At the then-peak of his career, Scott left San Francisco and moved back to his hometown in Louisiana to take care of his mother while she was terminally ill."
The public record of that period is thin, and I would rather leave it thin than fill it with invention. What can be said is this. Wu went home to Baton Rouge, to the house where the trophies stood in for photographs, and cared for his mother through lung cancer. She died in the autumn of 2023. A few weeks later—on November 17, the day the OpenAI board fired Sam Altman, the day every person in the industry he was about to enter was refreshing the same feeds—he founded Cognition.
The dates around the founding do not entirely agree. Some accounts place the start in August 2023; Wu's own LinkedIn and several company profiles give November. The likely reconciliation is the ordinary one: the work began earlier, in conversation and in code, and the company became a legal fact later. David Senra's account has the founders building it in a New York apartment. What matters is the conjunction. A son loses his mother; within weeks he has committed the next years of his life to the hardest problem he can find. One need not psychologize this to notice it. Wu told Senra, in a phrase the episode's own summary preserves, that he could accept failure but not regret.
The thesis was simple to state and, in late 2023, nearly universally doubted. GPT-4 had been public since March. The models could write code—snippets, functions, the kind of autocomplete GitHub Copilot had made familiar, the kind that reached 400,000 subscribers in its first month after public launch. What they could not do, in the consensus view, was engineer: hold a large codebase in mind, plan a sequence of changes, run them, read the failures, fix them, and deliver finished work without a human steering every step. That was years away. Everyone said so.
Wu looked at the same models and drew a different conclusion. "When it became clear in late 2023 that AI was about to be much more than an imitation engine," he wrote later, "and could actually think through problems the way humans do, a lot of us were really excited to work on something in the space." The economic logic, as he has laid it out, was almost insultingly direct: the world already spends enormous sums on human software engineers; the models could already write code; product-market fit was therefore not an open question but a foregone conclusion, provided you could make the thing work. "Being a good first principles thinker," he told the Z Fellows newsletter, "does beat experience."
He defined the job down to its bones. "What it means to be a software engineer in my mind is basically just somebody who solves problems in the context of code," he said. "It is somebody who tells the computer what to do and makes all these decisions." Stripped that bare, the problem stopped being a vague aspiration—automate software engineering—and became something with edges: a sequence of decisions under uncertainty, each one evaluable, the whole thing amenable to search. "Deeply algorithmic," he called it. It was, in other words, a contest problem. A very large one, with no clock and an unbounded prize pool, but a contest problem nonetheless.
Of everything we could have worked on, teaching AI to code was the most fun thing a group of programming nerds could imagine.
He assembled the group from the rooms. Steven Hao, the 2014 IOI medalist from Scale, became CTO. Walden Yan became chief product officer—a Harvard student who dropped out in 2023 to join, who had spent the summer of 2023, June through August, working on Cursor at Anysphere, the company that would become one of Cognition's fiercest competitors, and who had earlier co-founded a web3 venture called DeepReason. Yan was the youngest of the three and the one who had most recently seen the inside of a rival's product. Neal came too. So did Gennady Korotkevich, a name that in competitive-programming circles requires no gloss, and Andrew He. By March 2024 the team numbered ten and held, between them, ten IOI gold medals—a fact the company would cite so often it became a kind of corporate motto. Wu was frank that this was the point: teaching an AI to program, he argued, was a deep algorithmic problem requiring complex decisions, and nobody understood complex algorithmic decisions better than people who had spent their adolescence making them against a clock.
They pivoted eight times, by Wu's count, before landing on the approach that became Devin. The number is worth sitting with. Eight times in a few months a team of world champions decided that what they were building was wrong and threw it out—which is either a sign of poor initial judgment or, more plausibly, the contest instinct operating at company scale: abandon the dead branch, reallocate the time, try the next path in the tree.
Peter Thiel's Founders Fund put in $21 million in early 2024 at a $350 million valuation. Kim reports that Wu challenged Thiel to chess and poker as part of negotiating better terms. The detail is delivered as an anecdote of charm, and it is charming. It is also a little alarming, if you are the other side of the table. Most founders negotiate a term sheet. This one proposed to settle it on a board.
The Demo and the Debunking
Devin was announced on March 12, 2024, in a post on X, and the internet did what the internet does.
The announcement called it the first AI software engineer. It came with videos: Devin taking a task in plain English, producing a plan, opening a shell and an editor and a browser inside a sandboxed machine, writing code, running it, reading the errors, fixing them. Ashlee Vance, writing for Bloomberg under the headline "Gold-Medalist Coders Build an AI That Can Do Their Job for Them," reported watching it build a website in about ten minutes and recreate a playable Pong in a similar span. The company published a benchmark: on SWE-bench, a test built from real GitHub issues in real open-source projects, Devin resolved 13.86 percent end to end, unassisted. The prior best, GPT-4, had managed under 2 percent.
Thirteen point eight six percent is not, on its face, a world-changing number. It means Devin failed on six of every seven real-world problems it was given. But it was seven times better than the best general model, and the shape of the curve mattered more than its current height. Aravind Srinivas, the CEO of Perplexity, said it seemed to be the first demo of any agent, "leave alone coding, that seems to cross the threshold." Investors and engineers reposted it. Memes proliferated—a whole genre of short videos announcing the death of the software engineer. Wu's old MATHCOUNTS clip began to circulate alongside them, the boy who answered before the question was finished now the man who had built a machine to answer before the engineer was hired.
Then came the backlash, and it was not entirely unfair. Developers on social media condemned the product as a layoff engine at the precise moment, after two years of tech-industry cuts, when engineers were most sensitive to that prospect. Skeptics questioned whether it could handle anything with intricate requirements. A YouTube channel called Internet of Bugs went frame by frame through one of the launch demos—Devin purportedly completing a freelance job posted on Upwork—and argued that the agent had not actually done what the client asked, but had instead written, tested, and debugged code irrelevant to the request. The phrase "accused of lying" appeared in a trade-press headline.
Colossus, in its 2026 profile, put the matter with sympathetic bluntness: Devin "worked poorly, and he took intense public criticism for it."
What is striking, looking back, is that the criticism did not change the plan. The company turned down offers valuing it at $1 billion. In April 2024—The Information described Cognition as a "six-month-old AI coding startup"—Founders Fund led a $175 million investment at a $2 billion valuation. Investors were not buying the 13.86 percent. They were buying the trajectory and the team, the same bet a talent scout makes on a seventh grader who places first in the ninth-grade division. Wu, characteristically, framed it in terms of time horizon rather than vindication. In early 2026 he wrote: "When we launched Devin in 2024, we envisioned a future where every team had an infinite army of junior engineers. We were early and Devin wasn't good enough. Now it is!"
The general-availability launch came later in 2024, and Wu's own account of it, posted to LinkedIn in 2026, is the single most revealing document about how he runs a company.
Five days before Devin was due to become generally available, the team looked at the product and decided it was "close enough to ship but not yet what we wanted it to be." So that morning everyone was told to get on a flight to Park City, Utah, that afternoon, and to let family and friends know they would be unreachable for five days. All eighteen or nineteen of them moved into a single Airbnb. They barely left the house. They ran, Wu wrote, "on adrenaline and Diet Coke." Halfway through the week someone realized they ought to have a launch video, so they hired a local videographer, who filmed them in the house while they were still building the thing the video would advertise. Park City sits at roughly 7,000 feet. People started getting headaches. The company ordered a stack of oxygen tanks.
"There's this wild photo from that week," Wu wrote, "where around three or four in the morning the night before launch, everyone is plugged into oxygen and grinding. It's still one of my favorite memories."
— Scott Wu, on the Park City launch weekWe were just a group of people who genuinely liked one another, working on a product we actually thought was cool, and committed to giving it everything we had.
The comments under that post are an education in the cultural gap Wu operates across. May this level of utter abuse never touch my life again. Amen. This is supposed to be motivational? How could you be proud of this? And, from the other side: Forget the two-pizza rule. It's engineers per oxygen tank now. Both camps read the same photograph. One saw exploitation; the other saw the IOI, reconvened in a rented house with a bigger prize. What neither camp quite registered is that Wu's account is, above all, a story about finding his people. The oxygen tanks are incidental. The point, to him, is that everyone was in the room.
Seventy-Two Hours
On a Friday in July 2025, news broke that Google had struck a $2.4 billion deal with Windsurf, the agentic coding environment formerly known as Codeium. It was not an acquisition, exactly. It was what the industry had begun calling an "acquihire" or a licensing deal: Google would pay for a license to Windsurf's technology and hire its chief executive and a group of senior researchers, leaving the company itself—its product, its customers, its remaining several hundred employees—behind, a hull with no captain. Earlier in the year OpenAI had reportedly been in talks to buy Windsurf outright for considerably more. That deal had collapsed. Now the founders were gone to Mountain View and the rest of the ship was drifting.
"We heard the news that a deal was happening between Google and Windsurf on Friday, the same time as everyone else," Wu recalled to Stripe's Cheeky Pint podcast. "And we reached out to them cold that evening and started talking about it, and realised if there was something to do, it had to be ready to go by Monday. So we said, 'OK, if we want to explore this, we have to spend the entire weekend on this nonstop.'"
They reached a handshake agreement over that weekend. On July 14, 2025, Cognition announced a definitive agreement to acquire what remained of Windsurf: its product, its intellectual property, its brand, its enterprise customers, and its people. The company would later call it "the craziest 72 hours of our lives." Business Insider ran Wu's reaction in its headline: "I Live for These Moments."
The deal was a textbook application of the contest instinct to corporate strategy. Most acquirers would have needed weeks to evaluate a target, months to close. The market, for those 72 hours, was a board in flux—a valuable asset whose ownership was suddenly uncertain, whose employees were demoralized and newly available, whose customers were wondering who would answer the phone. Wu saw an opening that would exist for a weekend and not longer, and he moved inside it. Cognition, which had built an asynchronous agent that worked in the background, now owned an integrated development environment where engineers worked in the foreground. The two halves of the coding market—the tool you delegate to and the tool you type in—were suddenly under one roof.
The integration was not gentle. Within weeks, Wu sent a message to the combined workforce, later posted publicly, that would become a flashpoint in a wider argument about working hours in AI. Gergely Orosz, who writes The Pragmatic Engineer, reported that Cognition expected staff to put in more than eighty hours a week, and set the company alongside others adopting something close to China's "996" schedule—9 a.m. to 9 p.m., six days a week—a regime Chinese authorities had moved against in 2021. The Wall Street Journal followed two months later with a story about 100-hour workweeks in the AI arms race.
— Scott Wu, message to Cognition and Windsurf employees, 2025Cognition has an extreme performance culture, and we're upfront about this in hiring so there are no surprises later. We routinely are at the office through the weekend and do some of our best work late into the night. Many of us literally live where we work.
He went on to acknowledge that people who had joined Windsurf had not expected to join Cognition, and that while the company was proud of how it worked, "we understand it's not for everyone." The message offered buyouts to those who wanted to leave. It is, read carefully, an unusually candid document. Most companies with brutal cultures describe themselves in the euphemisms of "passion" and "ownership." Wu named the thing.
Some stayed. Jeff Wang, one of the Windsurf people who came across in the deal, wrote under Wu's post announcing the May 2026 funding round that working with him and "the most talent dense team in the world has been the experience of a lifetime." On June 2, 2026, Windsurf was rebranded as Devin Desktop—"the next generation of Windsurf," in the company's phrase. A product with its own name, its own brand, its own loyal following was absorbed into the name of the agent. The acquisition, eleven months on, had been made to speak in Devin's voice.
Switzerland
The commercial story of Cognition is, in its broad outline, the story of an agent that did not get good enough for enterprises until enterprises had already started paying for it.
The first meaningful traction came not from glamorous greenfield work but from drudgery. Nubank, the Brazilian digital bank, used Devin to upgrade Java versions across codebases of some 50,000 files—work in which, as Wu described it, the same handful of changes recur with minor variations thousands of times. It was the perfect first job for a junior engineer who never sleeps: too repetitive for a human to want, too varied for a simple script to handle. Wu built a theory of agent product-market fit around it. The fit appears first, he argued, in tasks repetitive enough to scope tightly but complex enough that automation without judgment breaks. Migrations. Modernizations. Refactors. Version upgrades. "It's crazy," he told Cheeky Pint, "how much of a software engineer's time is spent on things like fixing your Kubernetes deployments."
From there the company moved up the stack of difficulty and up the ladder of institutional caution. By 2026, Colossus reported, Devin was the agent of choice at enterprises ranging from the U.S. Army to Goldman Sachs to Mercedes-Benz. Cognition built a government business with FedRAMP High authorization in process. It opened offices in New York, Austin, Washington, London, Singapore, and Japan. The standard enterprise security and procurement cycle runs twelve to eighteen months; Wu claimed Cognition had compressed it to roughly three, by offering private-cloud deployments, strict data agreements, and forward-deployed engineers who sat with customers to find the highest-return work before handing it to the agent. "We show up," one of the company's four published principles reads. "We work alongside our customers—in real systems, under real constraints."
Two strategic decisions distinguished the company from its competitors, and both have the flavor of a game player thinking several moves ahead.
The first was model neutrality. Cognition does not train a frontier model of its own to sell; it builds the system around the models—a harness that routes each subtask to whichever model, from Anthropic or OpenAI or Google or the open-source world, offers the best combination of capability and cost. Wu calls this being Switzerland. A customer buying Devin, under this logic, is buying an agent whose loyalty is to the customer's outcome rather than to any one lab's tokens. In June 2026 the company released Devin Fusion, a hybrid harness that it said preserved frontier-level performance at 35 percent lower cost. Neutrality also has a defensive logic that Wu does not need to state: in a market where the model makers are also competitors, the company that depends on none of them is the hardest one to strangle.
The second was a refusal to sell activity. As AI coding tools proliferated, a habit spread through engineering organizations of measuring adoption by token consumption—how much model output each engineer burned through, as a proxy for how much they were using AI. Wu thought this was absurd and said so in a way designed to be quoted: "Measuring someone's productivity by their token usage is a horrible idea. Giving everyone the same fixed token budget isn't much better." Cognition built a system to estimate how many productive engineering hours each Devin task was worth, validated against real engineers' time estimates. Then it put money behind the estimate: an AI productivity guarantee under which, if Devin delivered less engineering value than a customer paid for, Cognition would fund the customer's usage until it did, up to $10 million.
It is a remarkable offer, and it rhymes with the poker game proposed to Peter Thiel. Wu is a person who prefers to settle disputes by outcome. If you think the product is overpriced, he will bet you it isn't.
The numbers, by any standard, were extraordinary. Devin's annualized revenue was $1 million in September 2024 and $73 million by June 2025. In April 2025, Devin 2.0 cut the entry price from $500 a month to $20, opening the product to individual developers. By its first eighteen months of service, Colossus reported, Devin had reached $445 million in revenue run rate, with usage doubling every eight weeks. On May 27, 2026, Cognition announced it had raised more than $1 billion at a $26 billion valuation, led by Lux Capital, General Catalyst, and 8VC—Lonsdale's firm, back again—with run-rate revenue of $492 million and enterprise usage up more than tenfold since the start of the year. By late summer, Bloomberg reported early talks at $40 billion or more; another outlet put a subsequent round at $47 billion. Forbes put Wu's net worth at $2.7 billion. The company acquired TierZero, a software-reliability startup, on July 20, 2026, and The Interaction Company of California, makers of an assistant called Poke, three days later. At some point that year, a reader who happened to be staying near Times Square posted under one of Wu's updates that he had seen Cognition's sign from his hotel window for two days running.
"There's about 30 to 35 million software engineers in the world today," Wu told Bloomberg Television. "We want to make all of them 10 times more efficient."
The Machine That Ships Its Own Code
In May 2025, Wu told Lenny Rachitsky that a team of Devins was producing 25 percent of Cognition's own pull requests, on track for half by year's end, and that each of the company's fifteen human engineers worked alongside about five of them. A year later the figure was 89 percent of the code committed internally. Other reports would put it higher still. "At this point," he said on one podcast, "our engineers don't type code anymore—you just tell the agent what you want it to do in English and it goes and builds."
This is the point at which the story turns strange, and Wu's own framing does a great deal of work to keep it from turning dark.
His preferred metaphor is architectural. Engineers will move from being bricklayers to being architects, he argues: "creatively structuring problems for armies of Devins to reliably execute on." He cites the Jevons paradox—the nineteenth-century observation that more efficient coal engines increased total coal consumption rather than reducing it—to argue that cheaper software will mean more demand for software, and therefore more engineers, not fewer. He talks about a world in which humanity shifts from "survival mode" to "creative mode," borrowing the vocabulary of Minecraft, the game in which survival mode makes you gather every block by hand and creative mode simply lets you build. The company's public principles insist that "the purpose of technology is to expand human capacity—not by replacing meaningful work but by working alongside people as an exponential collaborator."
He may well be right. The economics of Jevons have held for most general-purpose technologies. But the reader is entitled to notice what Wu is building and who he is. In the Invest Like the Best conversation in late 2024 he predicted that AI would surpass the world's best competitive programmer within one to two years. The world's best competitive programmers were, very nearly, his co-founders. A team holding ten IOI gold medals chose as its life's work the automation of the precise skill that had made each of them extraordinary—the thing that had given each of them, in childhoods spent as the unusual kid in some unremarkable town, a room full of peers. One commentator called Cognition "a cleaner signal about labor substitution than any forecast from outside the field." If the people who win at programming are building their own replacement, the rest of the profession's assumptions deserve revision.
Jeremy Stern, profiling Wu for Colossus in May 2026, opened his piece with a four-act sketch of the twentieth-century mathematical genius: virtuosity, persecution, madness, posthumous canonization—Nash, Turing, the whole tragic syllabus. His argument was that the arc had inverted. The nerd-genius no longer threatens the establishment; he is the establishment's most coveted asset.
— Jeremy Stern, Colossus, May 2026Today, those who share a field with the youthful nerd-genius no longer see in him a personal threat, but a power-law venture investment or nine-figure talent acquisition.
What Wu lacks, in Stern's telling and in every other account, is the old type's torment. There is no paranoia, no bathrobe in the seminar room. He solves card tricks and mental arithmetic for audiences and seems to enjoy it. He invests personally in other AI companies—Etched, Exa. He told Senra that money is not his motivation, and the claim is more credible from him than from most, because the evidence of his life is that the motivation has always been the same one: the next name called, and whether it is his. "Fear of losing," Senra's episode summarizes him as saying, "is a stronger motivator than winning."
There is also something gentler in the record, which the competitive frame tends to obscure. He flew to North Carolina to have dinner with the parents of a talented young recruit who were worried about their child joining an untested startup. When Kim lost someone in his family at the end of 2025, Wu was among the first to reach out, and Kim was given the space to step away immediately, no questions asked. The company's first written principle is not about winning. It is "The team is the company," and it asks for "kind, low-ego individuals." The man who described building a company as nothing but a tree search, who said competition was the only thing he does in his life, wrote of Park City that the best part was that everyone genuinely liked one another. The rooms, again. The people found by competing, kept by caring.
In September 2026, a Cognition engineer named Eric Lu announced that the company had factored RSA-260—a 260-digit number published decades ago as a public cryptographic challenge, the largest of its kind ever publicly broken—using a GPU implementation built and run with Devin. It was a feat of exactly the sort that would once have been the province of a prodigy alone at a blackboard. Lu signed the post. The machine did much of the work. Wu's name, so far as one can tell, appears nowhere on it.
He had told Senra, almost in passing, a small fact about the house in Baton Rouge: where other families keep photographs, the Wus kept the math trophies. It was the house he went home to in 2023, to sit with his mother through the end, before the November day he started building the thing that would do his work for him. The trophies are, by now, too many to count. None of them is from the second grade.
Part IIThe Playbook
What follows is an attempt to extract the operating logic beneath the story—not the biography of a prodigy, which is not reproducible, but the decisions, which mostly are. Wu has said that the conventional advice to founders—hire the best people, move fast, stay obsessively close to the customer, build something people want—is correct, and that the trick is not to look for secret truths but to apply the obvious ones to an extreme. Most of the principles below are, on inspection, ordinary truths taken further than most people are willing to take them.
The principles
- 1.Train on the loss.
- 2.Compress the mission until it becomes a problem.
- 3.Distrust pattern-matching on exponential curves.
- 4.Ship for the model you will have.
- 5.Recruit from the arena, and hire founders.
- 6.Go to North Carolina.
- 7.Move inside the 72-hour window.
- 8.Start in the drudgery.
- 9.Be Switzerland.
- 10.Sell outcomes, and bet on them.
- 11.Make your product your largest employee.
- 12.Name the price of intensity before anyone pays it.
⏱
The Compression Curve
From a second-grade loss to a $26 billion company, the intervals keep shortening.
2011
Wins MATHCOUNTS national championship as an eighth grader in Baton Rouge.
2014
First place overall at the IOI, his third straight gold; Steven Hao takes gold in sixth.
2016
Harvard team takes gold and third overall at the ICPC World Finals.
2017
Co-founds Lunchclub as CTO; serves until June 2022.
2023
Cares for his mother in Louisiana; founds Cognition in November, weeks after her death.
2024
Devin launches March 12; $21M at $350M, then $175M at $2B in April.
2025
$4B round led by 8VC in March; Windsurf acquired in July; $10.2B Series C in September.
2026
Over $1B raised at $26B in May with $492M run rate; Windsurf becomes Devin Desktop; RSA-260 factored with Devin.
Principle 1
Train on the loss
The formative event in Wu's competitive life was not a victory. It was a defeat in a competition he was five grades too young to enter, and the response was not consolation but a year of practice culminating in a perfect score two divisions up. The lesson is not "persevere," which is a poster. It is that the loss carries the most information. A win tells you that you were good enough on a given day. A loss tells you exactly where the gap is.
Wu's account of how to improve at competitive programming follows the same logic at small scale: read problems, solve some, give up on the hard ones, study them, return, repeat. The giving up is part of the method. A forum post attributed to him recommends looking up the solution after fifteen minutes stuck—not as a concession, but because the solution to a problem you have struggled with is the most efficient lesson available. At company scale, the eight pivots before Devin are the same move: treat each dead approach as data rather than as a verdict on the team.
The emotional component matters too. Wu describes being "salty" as an identity rather than a mood, and says fear of losing motivates him more than the pleasure of winning. Founders who treat competition as a temporary motivation tend to run out of it once the money arrives. Founders who treat it as who they are do not.
Tactic: After every lost deal, failed launch, or rejected pitch, write down the single specific gap it exposed, and schedule the work to close it before you schedule anything else.
Principle 2
Compress the mission until it becomes a problem
"Build an AI software engineer" is a mission statement, and mission statements are not solvable. Wu's move was to reduce the job to its irreducible core: an engineer is "somebody who solves problems in the context of code," who tells the computer what to do and makes the decisions along the way. Once defined that way, the work becomes a sequence of decisions under uncertainty—plan, act, observe, revise—which is to say a search problem, which is to say the kind of problem his team had spent their adolescence beating.
The compression did two things at once. It made the project tractable, because a search problem has components that can be built and measured independently: planning, tool use, testing, recovery from error. And it made the team's comparative advantage obvious, because "deeply algorithmic" problems are precisely where ten IOI gold medals pay off. A vaguer definition would have made the competitive-programming pedigree look like a hiring quirk. The precise definition made it the strategy.
The same habit runs through Wu's public thinking. "The whole point of software engineering is just to tell your computer what to do," he says, and from that sentence follows the entire product roadmap: whatever reduces the distance between intent and execution is progress, regardless of whether it looks like a traditional developer tool.
Tactic: Rewrite your company's mission as a single verb applied to a single object, and if you cannot identify which part of it your team is uniquely good at, keep rewriting until you can.
Principle 3
Distrust pattern-matching on exponential curves
Wu has argued that historical pattern-matching predicts the future correctly perhaps 99 percent of the time, and that AI is the 1 percent exception. In 2023, by the METR measurements he cites, models could complete perhaps ten to twenty seconds of uninterrupted human work; the figure doubled every few months until it was measured in hours. Anyone forecasting from the 2023 baseline, with the ordinary instinct that hard things take years, would have been off by orders of magnitude.
This is the logic behind founding Cognition when "almost nobody in the industry believed an AI software engineer was buildable yet." The consensus was not stupid; it was linear. Wu's edge was neither more information nor more experience—he had less of the latter than most skeptics—but a willingness to reason from the curve rather than from the precedent. "Being a good first principles thinker," he says, "does beat experience." That is a dangerous sentence in the wrong hands, and it is worth noting what makes it safe in his: he applies it only where he has evidence the underlying rate of change is unusual.
The test is not whether the consensus could be wrong; it always could. The test is whether the variable the consensus is implicitly holding constant is in fact moving fast. In AI capability it was. In most markets it is not.
Tactic: For any belief your industry holds about what is "a few years out," identify the single underlying capability it depends on, chart that capability's rate of improvement over the last 24 months, and extrapolate before you accept the timeline.
Principle 4
Ship for the model you will have
Devin, at launch, failed on six of every seven real-world tasks it was given, and Wu was publicly pilloried for overclaiming. He shipped anyway, turned down a $1 billion offer, and raised at $2 billion a month later. The decision only makes sense if you believe the product's job was not to be good in March 2024 but to be in position—in customers' Slack channels, in their Jira queues, in their procurement pipelines—when the underlying models became good enough to make it good.
"We were early and Devin wasn't good enough," he wrote in 2026. "Now it is!" The interval between those two sentences is where Cognition built its enterprise relationships, its deployment infrastructure, its evaluation systems, and its brand as the default name for the category. A competitor who waited for the models to mature would have arrived with a better first product and none of the surrounding apparatus.
The cost was real: reputational damage, the debunking videos, a reputation for hype that the company spent the next year working down. The principle is not to ship broken products indiscriminately. It is to recognize when the binding constraint on your product is someone else's rate of improvement, and to build everything that does not depend on that constraint while you wait for it to lift.
Tactic: List every component of your product that will still be needed when the core technology is ten times better—distribution, integrations, trust, evaluation—and build those now, even if the core is not yet ready.
Principle 5
Recruit from the arena, and hire founders
Cognition's founding team came almost entirely from the rooms where Wu found his people: Olympiad medalists, ICPC finalists, people he had competed against since adolescence. He advises working with people "you know well, like, and respect." That is a trust strategy as much as a talent strategy. When you have watched someone triage a five-hour contest under pressure, you know more about their judgment than any interview loop can reveal.
The second filter was entrepreneurial history. Of Cognition's first 35 employees, 21 had previously founded a company; at one point, 13 of 19. A team of former founders behaves differently from a team of strong employees. Each person has already made decisions with incomplete information and lived with the consequences. Each has a feel for urgency that cannot be taught. The obvious risk—too many captains—appears to have been offset by the shared competitive culture, in which status derives from solving the problem rather than owning the org chart.
⚖
Conventional Wisdom vs. the Wu Approach
Where Cognition took ordinary advice to an extreme, and where it inverted it.
| Conventional wisdom | Wu's approach | Mode |
|---|---|---|
| Hire experienced operators. | Hire former founders and competition peers whose judgment you have watched under a clock. | Extreme |
| Wait until the product is ready. | Ship when the trajectory is right; build distribution while the models catch up. | Inverted |
| Diligence an acquisition for months. | Reach a handshake in 72 hours while the board is in flux. | Inverted |
| Bet on one model provider. | Route every subtask to whichever model wins on cost and capability. | Structural |
| Measure AI adoption by usage. | Measure engineering hours delivered, and guarantee them up to $10M. | Bet |
| Soften the culture in hiring. | Describe the hours explicitly so there are no surprises. | Extreme |
Tactic: For your next five hires, ask each candidate to describe the hardest unsupervised decision they have ever owned, and weight that answer above their résumé.
Principle 6
Go to North Carolina
At one point Wu flew to North Carolina to have dinner with the parents of a talented young prospective hire, because the parents were worried about their child joining an unproven startup. It is a small anecdote, and it reveals a large principle: the highest-priority decisions deserve effort disproportionate to what anyone would reasonably expect.
The asymmetry is the point. A CEO's time is the scarcest resource in an early company, and spending a day on a single hire looks inefficient. But in a company where a handful of people determine the trajectory, the marginal exceptional hire is worth more than almost anything else the CEO could do that day. More subtly, the gesture communicates something to everyone who hears about it—that this company takes people seriously enough to show up for them. The same instinct runs through Cognition's customer motion, where forward-deployed engineers sit inside client organizations rather than handing over documentation, and through Kim's account of being given immediate leave during a family loss. "We show up" is written into the company's principles. The North Carolina dinner is what it looks like in practice.
Tactic: Identify the one person whose decision most affects your next twelve months, find the specific objection standing between you and them, and address it in person, even if it costs you a day.
Principle 7
Move inside the 72-hour window
When Google's licensing deal hollowed out Windsurf on a Friday, Cognition reached out cold that evening and concluded that if anything was going to happen it had to be ready by Monday. The team spent the weekend on it without stopping and emerged with a handshake agreement. The definitive deal was announced on July 14, 2025.
The general lesson is that some opportunities have a half-life measured in hours. A demoralized team with no leader will scatter within weeks; customers will begin evaluating alternatives within days; rival acquirers will mobilize within a week. The advantage went not to the company with the most money but to the one prepared to decide fastest. That preparedness was not improvised. A team that had already relocated to Park City on a morning's notice to finish a launch was a team for whom a weekend of nonstop negotiation was not an extraordinary demand.
The contest framing is useful here. In a timed competition, the value of a solution is not fixed; it decays with every minute. Wu said of the deal, "I live for these moments." The moments he means are the ones where the clock is the dominant variable and most competitors are still reading the problem.
Tactic: Pre-commit now to which kinds of sudden market dislocation you would act on within a weekend, and decide in advance who on your team would be in the room.
Principle 8
Start in the drudgery
Devin's first real enterprise success was not a glamorous demonstration of creativity; it was Java version upgrades across 50,000-file codebases at Nubank—the same small set of changes repeated thousands of times with slight variations. Wu's framework for agent product-market fit holds that it emerges first in tasks repetitive enough to scope tightly but complex enough that a simple script fails. Architecture design and novel debugging come later.
This sequencing is counterintuitive for a team of algorithmic virtuosos, who might naturally want to showcase the hard, interesting problems. It is also exactly right. Drudgery has three properties that make it ideal for an immature technology: humans resent doing it, so there is no resistance to delegation; it is easy to verify, so failure is cheap and visible; and its cost is easy to measure, so the return on investment is legible to buyers. "You reviewed the PR in five minutes, and that's all done," Wu has said. The value is self-evident.
▦
Where Agents Earn Trust First
Wu's sequencing of software work by readiness for delegation, drawn from Cognition's deployment experience.
| Task type | Why it fits or doesn't | Readiness |
|---|---|---|
| Migrations and version upgrades | Repetitive, scoped, easily verified by tests and CI. | Delegate |
| Bug reproduction and small fixes | Needs a working repository and an iterative test loop. | Delegate |
| New features from tickets | Works when completion criteria are explicit; drafts often need human takeover. | Pair |
| Broad debugging across systems | Requires understanding relationships, not just source text. | Pair |
| System architecture | The "architect" work Wu reserves for humans for now. | Human-led |
Tactic: Find the most repetitive, most resented, most easily verified task your customers perform, and make it the first thing your product does flawlessly.
Principle 9
Be Switzerland
Cognition does not sell any single lab's model. Its agent routes each subtask dynamically across Anthropic, OpenAI, Google, and open-source models according to cost and capability, a posture Wu describes as being Switzerland. The 2026 release of Devin Fusion, a hybrid harness the company said delivered frontier-level performance at 35 percent lower cost, is the commercial expression of the idea.
Neutrality solves three problems at once. For customers, it aligns Cognition's incentives with theirs: the agent will use whatever model produces the best result per dollar, not whichever one its owner wants to promote. For Cognition, it removes dependence on any single supplier in a market where suppliers frequently become competitors. And as models improve, the company captures each improvement the day it ships, regardless of who ships it. The risk is that the value accrues to the model makers and the harness becomes a commodity. Wu's wager is the opposite: that as models become interchangeable, the system around them—the planning, testing, environment, integrations, and enterprise trust—is where durable value lives.
Tactic: Audit every critical dependency in your product for a single-vendor chokepoint, and build an abstraction layer for any supplier that could plausibly become a competitor.
Principle 10
Sell outcomes, and bet on them
"Measuring someone's productivity by their token usage is a horrible idea," Wu wrote, and he built the alternative: a system estimating the productive engineering hours each Devin task represents, validated against real engineers' time estimates. Then he guaranteed it. If Devin delivers less engineering value than a customer pays for, Cognition funds the customer's usage until it does, up to $10 million.
The guarantee is a pricing move, a marketing move, and a forcing function. As pricing, it converts an uncertain purchase into a bounded one. As marketing, it reframes the industry's conversation from adoption to return, a framing that favors whoever can prove their returns. As a forcing function, it disciplines the company itself: a business that guarantees outcomes must measure them honestly, and a business that measures them honestly learns quickly where its product falls short. It is also, unmistakably, the same instinct that proposed settling a term sheet with Peter Thiel over chess. When there is disagreement about value, Wu's preference is to put something at stake and let the result decide.
Tactic: Define the single outcome metric your customer actually cares about, measure it independently of your own usage data, and offer a guarantee against it, even a small one.
Principle 11
Make your product your largest employee
In May 2025, Devins were producing a quarter of Cognition's pull requests, with each of fifteen human engineers working alongside about five. A year later, Devin shipped 89 percent of the code committed internally. Cognition is its own largest customer, and the arrangement is not symbolic.
Dogfooding at this intensity produces a feedback loop no external customer can provide. Every failure the agent makes on Cognition's own codebase is felt immediately by the people who can fix it. The company learns how teams must reorganize when agents do most of the typing—what Wu describes as the shift from bricklayers to architects—before its customers have to learn it, and can then teach them. It also lends the sales pitch a credibility that benchmark scores cannot. An engineering leader skeptical of autonomous agents can be told, truthfully, that the vendor bet its own product on them.
Tactic: Set a quarterly target for the share of your own internal work your product performs, and treat any month in which that share falls as a product failure.
Principle 12
Name the price of intensity before anyone pays it
Cognition's culture is, by most standards, punishing. The Park City launch ran on Diet Coke and supplemental oxygen at four in the morning. After the Windsurf acquisition, Wu wrote that the company had "an extreme performance culture," that its people routinely worked through weekends and late into the night, that many "literally live where we work"—and reports put expectations at more than eighty hours a week. Critics called it abuse. Many of the people inside call it the experience of a lifetime.
The defensible principle is not the hours themselves, which are a choice reasonable people will reject. It is the candor. Wu stated the terms openly, in writing, offered buyouts to Windsurf employees who had not signed up for them, and said plainly that the culture "is not for everyone." A culture that conceals its demands produces resentment and attrition; one that names them selects for people who want them. Pair this with Cognition's stated value of "go for it all"—"when faced with a tradeoff, pick the ambition-maximizing direction"—and the logic is coherent: a company optimizing for maximum ambition needs people who have consented, in advance and with full information, to what that costs.
Tactic: Write down the real working norms of your company, unvarnished, and put them in front of every candidate before the first interview rather than after the offer.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Scott Wu, to David Senra, June 2026Even building a company, it feels the same. It's just like, you're calculating the moves. You're thinking about, okay, if you do this, and then this happens, and then you do that, and here are the different moves, and you're calculating out what comes out to success.
— Scott Wu, Z Fellows, 2026The reason that young founders who are just really sharp and really determined did very well is because at the end of the day, being a good first principles thinker does beat experience.
— Scott Wu, Z Fellows, 2026What it means to be a software engineer in my mind is basically just somebody who solves problems in the context of code. It is somebody who tells the computer what to do and makes all these decisions.
— Scott Wu, LinkedIn, 2026Measuring someone's productivity by their token usage is a horrible idea. Giving everyone the same fixed token budget isn't much better.
— Scott Wu, LinkedIn, 2026At Cognition, one of our values is to go for it all: when faced with a tradeoff, pick the ambition-maximizing direction.
Maxims
- The loss is the lesson. A defeat tells you exactly where the gap is; a victory only tells you that you were good enough that day.
- Define the job before you automate it. An impossible mission becomes a solvable problem once you compress it to its irreducible core.
- Reason from the curve, not the precedent. Pattern-matching is right 99 percent of the time and catastrophically wrong on exponentials.
- Build the scaffolding while the engine catches up. If someone else's rate of progress is your constraint, spend the wait on everything that doesn't depend on it.
- Recruit from the arena. People whose judgment you have watched under a clock are worth more than people whose résumés you have read.
- Some windows close by Monday. When the board is in flux, the advantage goes to whoever is prepared to decide fastest.
- Begin with what people resent. The most tedious work is the easiest to delegate, the easiest to verify, and the easiest to price.
- Owe loyalty to the outcome, not the supplier. Neutrality aligns you with customers and insulates you from vendors who might become rivals.
- Put something at stake. When the value is disputed, a guarantee settles the argument faster than a pitch.
- Tell people the price up front. Intensity freely chosen is a culture; intensity concealed is a grievance in waiting.
