The Future of Software & AI | Cognition’s Scott Wu

Summary

• Senra recalls hearing that Cognition went from roughly $1M to ~$500M in revenue in about 20 months, but Wu does not confirm those figures; Wu says around 75–80% of current revenue is enterprise, including Goldman Sachs, Mercedes, parts of the US government and Nubank. Cognition tries to compress typical 12- to 18-month enterprise security-and-deployment cycles to within 3 months. Wu illustrates the value with organizations of 25,000 engineers costing something like $10 billion a year that believe they can move three times faster.
• Wu’s core thesis is that Devin could become—not merely a coding tool—the human–computer interface. Software engineering was always a way to tell computers what to do, and the abstraction keeps climbing from vacuum tubes and punch cards toward plain-language intent.
• The unpriced market, in Wu’s telling, is single-use software. Today software only pencils out if it will be used perhaps 10,000 to a million times; YouTube passes that test, but one-off white-collar tasks do not justify hiring engineers. Agents that write throwaway code on the back end could make one-time software economical, leaving humans to decide what to do.
• Senra cites a METR report showing uninterrupted AI task-length growing from 10–20 seconds to hours, with the capability roughly doubling every few months. Wu argues first principles beat historical pattern-matching in the rare periods when things truly change: “why can’t that be days or… weeks or months of work?” Seconds are commands, hours are tasks, years are missions—with a manager AI dispatching agents. His hedge: “I honestly think we’ll have solved most of that over the next 5 years or so.”
• Cognition’s deliberate neutrality—“we like being Switzerland”—is both a product architecture and a go-to-market stance. Devin is a compound AI system that dynamically routes subtasks across Anthropic, OpenAI, Google, open-source and Cognition’s own models. Cognition says it is not incentivized to make customers spend more on models; Wu calls token-spend ranking “directionally correct” but urges companies to measure output instead.
• Wu is betting on independence over acquisition, but says they would sell if it were the most ambitious path. He will not identify a price or the number of offers. His argument channels Spotify’s Daniel Ek—“we’re just going to care way more about music than they are”—and rejects the idea that GitHub Copilot or cloud incumbents automatically eliminate room for focused startups.
• The founder psychology on display is pure competition: “salty just means that you take offense to the idea of losing.” Company-building, to Wu, is “a tree search” through possible moves toward victory. Devin’s launch drew both “coolest thing ever” and “totally a scam” reactions; its SWE-bench result was 13% versus a prior best of roughly 3–4%, still failing 87% of the time but helping Cognition plant the AI-as-coworker flag.

Deep dive

1. “Salty” is the operating system: losing as offense, business as tree search

• Wu’s self-description opens the episode: his first competitive memory is entering a seventh-grade math competition at age seven or eight and being “so pissed” that his name was not called at awards. Wu defines salty as taking offense at the idea of losing.
• Company-building, in his framing, is the same game: “you’re calculating the moves… it’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 like the only thing I do in my life.”
• Senra compares Wu with Demis Hassabis of DeepMind: “they’re both really smart, they’re both articulate, they have like a friendly UI… but then underneath that is this ruthlessly competitive drive.” Senra recalls that Hassabis said—or, as Senra puts it, he thinks Hassabis said publicly—that half his brain is dedicated to competition.

2. A family that immigrated because of Go and hung trophies where photos should be

• Wu’s father was a seven-dan Go player. Wu tentatively compares that to roughly a 2,300–2,400 chess rating. Wu says his parents came to the US “in some sense” because of Go: a professor who played weekend Go with his father moved to the US, wrote back about the opportunities and helped him apply to graduate school.
• His mother was “definitely the saltiest” in the family. Senra’s biographical taxonomy—the Ellison/Musk fathers who said “you’re worthless” versus the Sara Blakely/Estée Lauder-style encouragers—places her in the latter camp: she told Wu he was the best even when he was tiny, and Wu thinks that was before there was evidence.
• Wu qualifies that his parents would have been happy if he had taken a more traditional, cushy job; they did not specifically steer him toward entrepreneurship. But his mother was a huge source of support and always believed in what he wanted to do.
• The household iconography was trophies: “we didn’t have pictures of our parents on the walls, we just had old math competition trophies.” Competition circuits, not Baton Rouge, supplied Wu’s closest childhood friendships. He also played Super Smash Bros. Melee in tournaments, Tetris, poker, some chess—“okay,” but not good—and some Go.

3. Losing hurts more than winning feels good—but not enough to stop

• Senra brings the Larry Ellison line—“I’m addicted to winning, but I fear losing more than I love winning”—which Michael Dell confirmed to him directly. Wu’s useful qualification is that the great entrepreneurs got where they are “by losing a lot”; the pain cannot be allowed to stop the attempts.
• His formulation is worth keeping verbatim: “losing feels way worse than winning feels good, but not by enough that it makes me want to stop trying.”

4. The endgame is broader than a coding agent: it is how humans tell computers what to do

• Cognition’s nine-person founding team was mostly made up of people who had already founded companies. They converged on one idea: “this is the big one… we want to be a generational business… we want to build a hyperscaler. Maybe we’ll succeed, maybe we won’t, I don’t know, but that’s what we’re going after.”
• Wu’s broader reframe is that “we’re not going to be interacting with code for that much longer,” while what remains true is the human–computer interface. Software engineering was a means of telling a computer what to do; Devin, if Cognition succeeds, becomes that interface for the world.
• The civilizational version, via his co-founder’s line: “we’ve been spending all this time living in survival mode as a species, and now we’re going to be living in creative mode.” In Minecraft creative mode, resources are available and “the only question for you is what you want to make happen.”

5. The hidden market: software that’s only used once

• Devin today is an AI software engineer used by Goldman Sachs, Mercedes and “a lot of areas of the US government.” Wu says teams use it to ship 10 times faster and do 10 times more. His macro claim is that “software is eating the world” still has a couple of orders of magnitude to go.
• The economic mechanism he spells out is qualified: software today only makes sense “if it’s going to be used at least like a million times or something”—perhaps 10,000 times—because engineers are expensive. YouTube passes; the one-off tasks of white-collar life do not justify a team, so people do them manually.
• Agents invert that math. You describe the intent, and on the back end “the agent is going to figure out, okay, I’m going to write this code… put a script that automates this part.” That could make single-use software economical while leaving the human responsible for deciding what to do.

6. First principles beat pattern-matching in the 1% of times things are actually different

• Senra frames the problem through exponential change: historical pattern-matching works “99% of the time,” but the rare periods when the world genuinely changes require first-principles thinking.
• Senra cites the METR report’s example of AI doing 10–20 seconds of uninterrupted human work a couple of years earlier, with that capability roughly doubling every few months and now reaching hours. Wu takes the next step: “why can’t that be days or… weeks or months of work?”
• Wu says humans lack an intuitive signal for these scales: a good hunt might produce a couple days of food, while an equivalent exponential gain could represent 1,000 years’ worth.
• His parents’ experience moving from 1960s communist China to the US is his lived example of jarring change. People adapted to cars and household appliances quickly; he expects a similar process with AI, saying that in 10 years “we’re going to have forgotten that we ever lived without them.”
• His timeline remains hedged: “I honestly think we’ll have solved most of that over the next 5 years or so.”

7. From bulk email to missions: what a year-long agent is for

• Wu’s frustration the day before—fighting an email formatter that would not let him unindent one part—crystallizes the gap: “it’s really crazy that I’m still doing this… the rest of that execution should just be done for you.”
• Senra’s pushback lands: “I don’t want your year-long agent to send bulk email.” His counter-example is Edwin Land hiring someone at Polaroid to solve the problem of turning instant photography from black-and-white into color; the person reportedly spent two years thinking before solving it.
• Wu agrees and draws the ladder: seconds are commands, hours are tasks, years are missions. Examples include a societal problem, a video game combining elements of two others, or a novel construction of materials that an agent studies, explores and tests for months or years. Senra would have one agent choose the missions; Wu calls that “the manager AI of the missions.”
• The ancestors analogy shows how work itself may be redefined: “you’re just pushing buttons… you call that a meeting and that’s work for you guys?” Future humans pursuing curiosities through dispatched agents may think of that as work, while people today will find the idea startling.

8. The launch: 13% on SWE-bench, hate, and a sleepless night over MongoDB

• Wu says two things were constant from the beginning in 2023: Cognition would focus on code, and its systems would handle “real, multistep, iterative processes”—a hot take when the dominant experience was closer to chatbot Q&A and autocomplete.
• Devin’s first real task was setting up MongoDB. After the team repeatedly encountered errors, they told it to run the commands and fix the problem. It worked. “I could not sleep that night,” Wu says; the team saw a better-than-average run and wondered, “why shouldn’t all software and all products be built this way now?”
• The March 2024 demo drew both ends of the spectrum: “the coolest thing ever” versus “there’s no way this is ever going to work, this is totally a scam.” Its SWE-bench result was 13% when the best known result was roughly 3–4%—still an 87% failure rate. Wu says the criticism did not push them toward a chatbot-style product, though he allows they may have ignored that possibility more than they should have.
• On being late—OpenAI began in 2015, DeepMind and Google Brain were already more than a decade old, and GitHub Copilot had been used for years—Wu says, “when you start a company… you have no right to exist.” The escape is to plant a flag, run toward a future in which AI is a co-worker rather than merely a tool, and adjust the details along the way. “We were wrong on a lot of things,” he says.

9. Enterprise product-market fit: migrations and flying the whole company to Brazil for Nubank

• The post-launch pilots in April and May 2024, built with GPT-4-era agents, “were all just failing”: the system could produce cool toy demos but was not ready for real companies and codebases. The strategic question became: what is the first task where an end-to-end agent can deliver value?
• The answer was work repetitive and scoped enough for a tight feedback loop but requiring “some intelligence and some amount of meandering.” Initial use cases included migrations and version upgrades, such as Java 7 to Java 8, across a 50,000-file codebase requiring the same eight changes in each file.
• The first success was Nubank, described as the biggest bank in Brazil by market capitalization at the time, using a custom Devin optimized for the migration. “Our entire team was the forward-deployed team—we all flew to Brazil… let’s be real here, agents did not generally work.” Senra’s summary: “you didn’t deploy a team, you deployed the whole company.”
• Today, around 75–80% of revenue is enterprise, while self-serve has also grown, with users including Exa, OpenRouter and Bilt. Cognition tries to compress the usual 12- to 18-month enterprise adoption cycle to deployment within 3 months, using private-cloud deployment, strict data agreements and security controls.
• The current forward-deployed motion is mostly education and guidance: identifying the right use cases, helping with setup and playbooks, and showing customers where Devin can and cannot create value. Customers, rather than Cognition employees, run the work.

10. Incentive alignment and the Switzerland bet

• Wu calls token-spend culture “directionally correct” but says people have gotten carried away—for example, ranking engineers by tokens rather than by the output and useful work they produce. The math can work: if engineers ship three times more, expensive GPUs may still be clearly worth it.
• The concrete ROI pitch is a project scoped for 18 months, destined for an outsourced contractor at $15 million, that might instead be completed internally for $1 million in 3 months. Cognition also tells customers which projects Devin cannot make 10 times faster and what workflows to use instead.
• The Switzerland architecture is Devin as a compound AI system. A bug ticket may require reading the report, reproducing the bug locally, inspecting logs, pinpointing files, debugging, retesting and opening a review. Different models are suited to different subtasks: the strongest models for difficult reasoning, and cheaper, faster, easily verifiable models for boilerplate.
• Devin can route among models from Anthropic, OpenAI, Google, open source and Cognition itself. Wu says Cognition is incentivized to make customers’ model spend efficient, not to make them spend more.

11. Independence: “we would sell if we thought it was the most ambitious thing to do”

• The focus doctrine, via Daniel Ek on Spotify versus Apple Music, is: “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.” Cognition’s version is caring deeply about the messy reality of building software end to end at Goldman Sachs or Mercedes—codebases, ticketing systems, testing, collaboration and deployment—not just solving a sandbox algorithms problem.
• Senra recounts Jimmy Iovine’s Spotify-versus-Apple framing: Apple’s resources initially looked like an asset but became a liability, while Spotify cared more intensely about winning. Wu heard a similar argument that Microsoft and GitHub Copilot already owned the coding market. He calls that an “uncreative” way to think about startups and invokes Datadog and Snowflake as examples of focused companies that existed despite cloud incumbents.
• On selling, Wu will not say how many offers Cognition has received—“dozens is a lot, dude”—and Senra recalls asking him about an acquisition by an unnamed company, to which Wu replied, “I don’t know how we’d be able to afford them.” Wu’s actual position: “we would sell if we thought it was the most ambitious thing to do. It’s kind of an oxymoron.” Lifestyle evidence: he has no car, rents an apartment and likes eating sushi.
• Senra rejects the idea that the current Devin interface is the permanent endpoint: there may be 10 more generations of product experiences to build. He also invokes Richard Branson’s definition of business—“an idea that makes somebody else’s life better”—to argue that the opportunity is not exhausted; there are indefinitely many ways to improve people’s lives.
• The closing logic is self-admittedly circular. Failing after giving it everything “would be an outcome I could live with—I’d be salty as hell”—but failing to push as hard as possible would be unbearable. Senra notes that Wu is the second-wealthiest entrepreneur from Baton Rouge, behind Raising Cane’s founder Todd Graves, who has also turned down major acquisition offers. Wu’s conclusion: “We want to achieve our potential and build what we were meant to build… you have one life, go.”