David SenraBuilding The World's First AI Software Engineer | Cognition’s Scott Wu
CHAPTERS
- 0:02 – 2:04
A lifelong obsession with winning (and why Scott calls himself “salty”)
Scott explains that his earliest memories are defined by intense competitiveness and a visceral dislike of losing. He connects that drive to how he thinks about company-building as a strategic game: calculating moves and searching decision trees for the path to victory.
- •Early childhood story: competing in an older kids’ math contest and being furious not to place
- •Competition as an identity: "it’s all I do"
- •Building a company feels like strategy games and tree-searching through decisions
- •The "friendly UI" masking a ruthless competitive core
- 2:04 – 3:06
Competitive programming as a tribe: finding “his people” through escalating contests
Scott describes how math and programming competitions created a ladder of increasingly challenging environments—and a social world of peers who were more like him than classmates at home. He also recounts the many competitive games he played and how they reinforced his mindset.
- •Goal as a kid: become world champion in competitive programming
- •Competition structure (school → city → regional → national → international) creates community
- •Online friendships formed around problem-solving and shared obsession
- •Other competitive arenas: Smash Melee tournaments, Tetris, poker, some chess
- 3:06 – 8:31
Family roots of competitiveness: Go, immigration, and a mother who instilled confidence
Scott ties his family story to competition: his father’s high-level Go shaped both the household culture and, indirectly, the family’s move to the U.S. He credits his mother’s pride and pre-evidence belief in his talent for building self-confidence and a home that celebrated achievement.
- •Father as elite Go player; Go relationship helped open a path to U.S. grad school
- •Contrast of parents: mom as the “most salty,” competitive by personality
- •Mother’s confidence-building: told him he was the best, even before proof
- •Family value system symbolized by trophy displays at home
- 8:31 – 9:38
Why losing hurts more than winning feels good (and why that doesn’t stop him)
Scott agrees with the idea that fear/pain of losing can exceed the joy of winning, echoing founders David cites. He emphasizes that progress requires frequent losses and repeated exposure to failure, and that the discomfort isn’t strong enough to deter continued striving.
- •Affirms: losing feels worse than winning feels good
- •Winning requires losing often; repeated attempts are unavoidable
- •Pain doesn’t lead to quitting—drive overrides aversion
- •David frames it as a personality type that can’t not compete
- 9:38 – 12:55
What “winning with Devin” means: building the human–computer interface
Shifting to Cognition, Scott defines victory not as a feature set but as building a generational company that becomes a foundational layer. He argues Devin is ultimately about enabling humans to tell computers what to do—an evolving interface that may outgrow today’s code-centric abstraction.
- •Founding team’s ambition: “this is the big one,” aiming for a generational hyperscaler
- •Devin as more than a coding agent: the future human–computer interface
- •Abstraction will keep rising; code may not remain the main level of interaction
- •Motivation: giving people the power to create and express ideas through software
- 12:55 – 13:57
Devin today: an “AI software engineer” used by major enterprises to ship faster
Scott outlines Devin’s current positioning and how companies use it end-to-end for software work. He lists major customers and claims teams use Devin to dramatically increase throughput and shipping velocity.
- •Devin described as an AI software engineer that collaborates end-to-end
- •Used by large orgs (e.g., Goldman Sachs, Mercedes, U.S. government areas)
- •Promise: ship far more work with existing teams (e.g., “10x faster”)
- •Context: software continues to expand across every industry
- 13:57 – 18:45
Software as evolving abstraction: from punch cards to one-off automation for everyone
Scott gives a history-of-programming view to argue the interface will continue to climb in abstraction. He highlights a key unlock: once agents make software cheap enough, people can create one-time or few-time-use software to automate work that isn’t worth building today.
- •Programming is fundamentally "telling computers what to do"; interfaces keep changing
- •Near-term reality: you don’t need to know languages to request software changes
- •Economic shift: software currently must be reused many times to justify engineering cost
- •Agents enable one-off automation for white-collar tasks that are currently manual
- 18:45 – 20:39
Why AI progress is hard to intuit: exponentials, culture shock, and the next 5 years
Scott argues people systematically underestimate AI because human intuition isn’t built for exponential curves. He predicts rapid change on a short timeline and draws on his parents’ experience of dramatic societal shifts as a reference point for how quickly humans adapt to new baselines.
- •Humans misread exponentials; AI scaling laws and company curves reflect this
- •Prediction: major parts of the vision solved within ~5 years
- •Analogy: parents’ move from 1960s China to the U.S. showed how fast life can transform
- •Expectation: we’ll soon forget what it was like to live without AI capabilities
- 20:39 – 22:57
Thinking from first principles: from seconds of autonomy to months (and then missions)
Scott says the right way to forecast is to avoid historical pattern-matching and instead reason from first principles about autonomy horizons. If AI can do hours of uninterrupted work today, he asks why that won’t become days, weeks, or months—implying a fundamentally different world of delegated labor.
- •Critique of pattern matching; first-principles reasoning matters in discontinuities
- •Autonomy benchmark framing: from ~10–20 seconds to hours of work without interruption
- •Core question: why not days/weeks/months?
- •Implication: widespread access to agents doing long-horizon work changes everything
- 22:57 – 30:07
Year-long agents and “missions”: from tedious tasks to deep problem-solving and discovery
Pressed on what long-horizon agents should do, Scott reframes usage from tasks to missions—multi-month or multi-year pursuits. He gives examples spanning societal coordination, creative work like games, and scientific exploration, and argues future “work” will look alien compared to today’s meetings and tools.
- •Long autonomy enables mission-based delegation, not just task completion
- •Examples: advocacy/coordination, designing novel games, pursuing scientific/materials ideas
- •Contrast with today’s friction (e.g., email formatting) as a symptom of poor interfaces
- •Historical perspective: future work will seem as strange as modern desk work would to ancestors
- 30:07 – 31:07
Cognition’s original thesis: code-focused agents doing multi-step iterative processes
Scott explains Cognition’s founding premise: stay centered on software and build agents that can execute iterative, multi-step workflows—not just Q&A chat. He positions this as a contrarian take in late 2023 and connects it to the team’s background as programmers and prior founders.
- •Two-part thesis: code/software focus + real multi-step iterative agent workflows
- •Timing: late 2023/early 2024, after ChatGPT; entering a crowded ecosystem
- •Founder-team profile: many had built companies before; large founding team
- •Vision: agent as coworker rather than autocomplete/chatbot tool
- 31:07 – 37:28
Launching Devin, handling criticism, and the early “prototype” reality
Scott recounts the polar reactions to Devin’s viral launch and admits it was closer to a prototype than a product. He describes the first “holy shit” moment when Devin solved a real setup problem, and why the team kept believing in the trajectory despite public skepticism.
- •Launch reactions ranged from “jobs gone tomorrow” to “it’s a scam”
- •Early system was a demo/prototype showcasing real internal runs
- •First breakthrough: Devin debugging MongoDB setup end-to-end; Scott couldn’t sleep
- •Belief: capability growth would continue exponentially regardless of early criticism
- 37:28 – 42:41
Finding enterprise PMF: start with repetitive, scoped work (migrations) and win with outcomes
After early pilots failed on messy real codebases, the team identified the first viable wedge: repetitive but non-trivial enterprise tasks with tight feedback loops. Scott details the Nubank migration as the first major success and explains why large companies became a natural target market.
- •Early pilots: many failures using GPT-4-era agents on real repos
- •Wedge selection: repetitive/tedious tasks needing intelligence (e.g., upgrades/migrations)
- •First big win: Nubank migration with a heavily optimized/custom Devin
- •Thesis: biggest companies are software companies; focus on real production environments
- 42:41 – 48:18
Deploying Devin in large companies: onboarding, security, and a “forward deployed” playbook
Scott walks through enterprise adoption: education first, then navigating long procurement and security cycles, with Cognition compressing timelines. He shares an extreme early story—flying the whole team to Brazil for Nubank—and contrasts it with today’s more scalable deployment and enablement motion.
- •Enterprise motion: educate on agents, guide to feasible use cases, set ROI expectations
- •Security/procurement can be 12–18 months; Cognition targets deployment in ~3 months
- •Early reality: entire team flew to Nubank to debug and tailor workflows on-site
- •Today: forward-deployed help focuses on enablement, playbooks, setup—not doing work for them
- 48:18 – 57:02
Measuring ROI over token spend: model-neutral “Switzerland” and focus as a startup advantage
Scott argues customers should optimize for output and business outcomes rather than token metrics. He explains Devin as a compound-model system that dynamically picks different models for different subtasks, enabling cost-performance optimization—and says focus is how startups beat giants despite incumbents’ resources.
- •Pushback on token obsession: measure shipped output and business impact instead
- •ROI framing: tie Devin to specific projects, timelines, and cost reductions
- •Model neutrality: Devin as “Switzerland,” routing subtasks to best-fit models (including their own)
- •Startup edge: extreme focus on real-world software workflows; "care more" than big platforms
- 57:02 – 1:05:10
Independence, acquisitions, and the one-life mindset: going for a generational company
Scott declines to quantify acquisition interest but emphasizes the ambition to remain independent if that’s the most ambitious path. He downplays money as a motivator and frames success as fully pursuing their potential—accepting failure is tolerable, but not trying hard enough isn’t.
- •Independence as a deliberate stance amid high-profile AI acquisitions
- •Selling only if it’s the most ambitious option; money not the North Star
- •Founder psychology: rejecting “it’s too late” narratives; many product generations remain
- •Closing ethic: one life—go for it all; regret comes from not pushing to the limit
