Scott Wu is the co-founder and CEO of Cognition, the company that built Devin, an AI software engineer used by enterprises like Goldman Sachs, Mercedes, and parts of the US government. The episode traces his obsessive competitiveness from childhood through competitive programming, his founding of Cognition, and his vision of AI as the future human-computer interface.
Scott Wu’s Obsession With Winning
Wu describes himself as “salty” — someone who takes deep offense at losing — and says this drive has defined him since childhood.
His earliest competitive memory: at age seven or eight, he entered a middle school math competition, expected to place, didn’t get called, and was furious.
He frames all of life, including company-building, as a tree search: calculating moves and counter-moves to reach victory.
He compares himself to Demis Hassabis (DeepMind), noting both have a friendly exterior but a ruthlessly competitive core.
Competitive Roots: Math, Programming, Games, and Family
Wu’s childhood centered on math and programming competitions, which gave him a ladder (school → city → regional → national → international) and introduced him to his closest friends.
He competed broadly: Super Smash Bros. (Melee) tournaments, Tetris, poker, chess, and Go.
His father was a strong Go player (roughly 7-dan), which indirectly led to the family immigrating to the US when a professor father played with moved there and encouraged him to follow.
Wu describes his mother as the most competitive (“salty”) person in the family.
She told him he was the best even before there was evidence, hung math competition trophies on the mantelpiece instead of family photos, and intentionally displayed other people’s trophies as a challenge for Scott and his brother to replace them with their own.
Why Losing Hurts More Than Winning Feels Good
Wu agrees with a pattern he sees in founders like Larry Ellison and Michael Dell: the pain of losing is worse than the thrill of winning.
But the gap isn’t large enough to make him stop trying.
He believes you have to lose a lot to get anywhere, and putting yourself out there is the only path.
Why Cognition Exists: Teaching AI To Build Things
Cognition’s founding team was nine people, most of whom had already started companies before. They came together believing “this is the big one.”
Wu’s core idea: software engineering is really just the human-computer interface — the way humans tell computers what to do.
If successful, Devin becomes the way anyone tells a computer what to do, not just engineers writing code.
He believes this is a generational opportunity: “doing that for the world is a massive opportunity.”
His deeper motivation is the shift from “survival mode” to “creative mode” for humanity.
He borrows a Minecraft metaphor from co-founder Steven Cao: moving from survival mode (securing food, shelter, safety) to creative mode (having all resources at your disposal and deciding what you want to make happen).
He sees the most human thing as self-expression and creativity — having things you want to make happen and being able to go do them.
Devin Today and the Path Forward
Today, Devin is an AI software engineer that works end-to-end with software teams at major companies, helping them ship roughly 10x faster and do 10x more.
Wu sees the abstraction climbing: you already don’t need to know Python or Java to build software, and eventually you won’t need to look at code at all.
He distinguishes between software used millions of times (where today’s model works) and tasks done once or a few times (where it doesn’t make sense to hire engineers today but where an agent could write a one-off script on the fly).
He believes most of this future will be solved in the next 5 years — “in AI terms, 5 years is like a century.”
He notes humans are poorly wired to intuit exponential curves, which is why people consistently underestimate how fast AI progress will change things.
First-Principles Thinking About AI Progress
Wu says most people predict the future by pattern-matching on the past, which works 99% of the time but fails during real discontinuities.
The better approach is first-principles thinking: asking why an AI that can do hours of unassisted work couldn’t eventually do days, weeks, or months of work.
He references the METR observation that AI’s ability to work without interruption went from ~10–20 seconds to hours in roughly a year, doubling every few months.
If that continues, agents working for months or years unassisted becomes plausible, and that changes what the world looks like.
What He Would Do With a Year-Long Agent
Wu distinguishes levels of delegation: seconds = a command, hours = a task, years = a mission.
He doesn’t want a year-long agent formatting emails; he wants it given a mission.
Examples: solving a societal problem he cares about, designing a dream game by combining elements from games he loves, or exploring a novel approach to materials science.
He imagines people sending agents on missions while they focus on deciding what missions to pursue — a manager-of-agents model.
He references Edwin Land (Polaroid founder), who hired someone to think for two years about how to turn instant photography from black-and-white to color; Wu wants that kind of sustained, focused agent effort on demand.
The Original Thesis Behind Cognition
Cognition’s thesis from the start (2023): work with code/software, and focus on real multi-step, iterative processes — which was a hot take at the time.
Wu recalls hearing about the early team scaling from no revenue to ~$500 million in roughly 20 months and thinking the idea of an automated software engineer was huge because it was going after labor.
Launching Devin and Handling Criticism
The March 2024 launch was a demo/prototype, not a real product, but it went viral.
Reception was polarized: some called it the coolest thing ever and predicted mass job loss; others called it a scam that would never work.
Wu describes the first time Devin completed a real task (setting up MongoDB, debugging through errors autonomously): “I could not sleep that night” because he saw the exponential curve ahead.
Even though it was a single best-in-class run, the feeling was: “Why shouldn’t all software be built this way now?”
At launch, Devin solved ~13% of SWE-bench tasks, versus ~3–4% for the best-known systems at the time — still failing 87% of the time, but a meaningful leap.
Wu argues Cognition was actually late (OpenAI, Google DeepMind, GitHub Copilot all predated them) and had “no right to exist” on paper, but won by planting a flag on the idea of AI as a co-worker rather than a chatbot tool.
Finding Product-Market Fit in the Enterprise
After the viral launch, early pilots with companies were failing because Devin wasn’t ready for real codebases.
Wu’s team asked: what’s the first task that gets real PMF? Answer: repetitive, tedious, scoped tasks with tight feedback loops — not open-ended architecture work.
This led to use cases like migrations and version upgrades (e.g., Java 7 → Java 8 across 50,000-file codebases).
First real success: Nubank (Brazil’s largest bank by market cap at the time), using a custom Devin optimized for a large migration.
The entire Cognition team flew to Brazil and essentially built the product alongside Nubank’s engineers.
Today, 75–80% of revenue comes from enterprise; the rest from startups and growing self-serve teams, all of them real engineering teams building real products.
How Cognition Deploys Devin Inside Large Companies
Enterprise onboarding normally takes 12–18 months; Cognition tries to do it in under 3 months.
They deploy in customers’ private clouds, with strict data agreements and tight air-gaps.
The motion is now more about education, guidance, and pointing teams to the right use cases than physically embedding.
Wu emphasizes incentive alignment: Cognition tells customers which projects Devin can make 10x faster and which it can’t, rather than just handing over a tool.
Measuring ROI Instead of Token Spend
Wu thinks the industry focus on token spend has gotten a little crazy (e.g., ranking engineers by tokens used).
He prefers measuring actual output: tickets closed, projects shipped, cost and time saved.
Example: a project scoped for 18 months and $15 million via outsourced contractors done internally for $1 million in 3 months.
Why Cognition Wants To Be Model-Neutral
Wu describes Cognition as “Switzerland” regarding AI models.
Devin is a compound model system: it dynamically chooses among models from Anthropic, OpenAI, Google, open-source models, and Cognition’s own models depending on the task or subtask.
Hard architecture problems get the strongest, most expensive models; repetitive boilerplate gets fast, cheap, verifiable models.
This aligns incentives: Cognition is incentivized to optimize price-performance and customer value, not to push more token spend.
Why Focus Lets Startups Beat Giants
Wu believes startups win by making a concentrated bet on one thing and caring about it more than incumbents do.
He cites Daniel Ek’s answer to why Spotify should exist despite Apple Music and YouTube: “We’re just going to care way more about music than they are.”
For Cognition, that means caring more about the messy reality of building software at Goldman Sachs or Mercedes-Benz than general-purpose labs do — plugging into ticketing systems, testing code locally, understanding existing codebases, collaborating with human engineers.
He pushes back on the “it’s too late / the labs will do everything” narrative as uncreative, noting that every era has the same incumbents-versus-startups dynamic and startups keep finding niches.
Independence, Acquisitions, and Building a Generational Company
Wu confirms Cognition has received acquisition interest but declines to say how many or from whom.
He jokes that when a mutual friend raised the idea of Cognition being acquired by a major company, his first reaction was “I don’t know how we’d be able to afford them” — assuming he’d be the buyer.
The founding team is committed to building a generational, independent business.
He pushes back on “nihilism” about whether new independent companies can still be built, calling founders “rationally optimistic.”
Why Money Is Not the Goal
Wu says he doesn’t have a car, rents apartments, and his main indulgence is sushi — which an engineer salary could cover.
He would only sell Cognition if selling were the most ambitious thing they could do.
He frames it as wanting to achieve their potential and build what they were meant to build: “if we felt like we could have gone for it all and we didn’t… I don’t think we would live with ourselves.”
He identifies with founders like Steve Jobs or Todd Graves (Raising Cane’s) who turned down life-changing money because they wanted to keep building.
