Scott Wu, Cognition

What It Covers

Scott Wu, CEO of Cognition and creator of AI software engineer Devon, discusses building a generational AI company from scratch in early 2024, scaling from zero to $500M+ revenue in roughly 20 months, targeting Fortune 500 enterprises, and his vision of AI agents eventually replacing all manual software execution within five years.

Key Questions Answered

• •Competitive identity as strategic asset: Wu traces his ruthless competitive drive to childhood math and programming competitions, where he advanced from local to international levels. He frames company-building identically to competitive gaming — a decision tree search calculating optimal moves toward victory. Founders who internalize competition as identity, rather than motivation, sustain intensity through multi-year execution without external incentives like acquisition offers or financial milestones.
• •First-principles thinking over pattern matching: Wu argues that 99% of the time, historical pattern matching predicts the future accurately — but AI is the 1% exception. The METR benchmark showed AI completing 10-20 seconds of uninterrupted human work in 2023; that figure has doubled every few months, reaching hours. Founders and investors who apply standard pattern matching to exponential curves systematically underestimate AI's trajectory by orders of magnitude.
• •Enterprise go-to-market: compress 18-month cycles to 3 months: Devon's enterprise sales motion targets Fortune 500 software teams — Goldman Sachs, Mercedes, US government — where software engineering organizations cost billions annually. The standard enterprise security and procurement cycle runs 12-18 months. Cognition compresses this to roughly 3 months by prioritizing private cloud deployment, strict data agreements, and forward-deployed teams who guide customers toward high-ROI use cases before handing off execution.
• •Start with repetitive, scoped tasks to prove agent PMF: Devon's first enterprise success came from code migration work at Nubank — upgrading Java versions across 50,000-file codebases, where the same 8 changes repeat with minor variations. Wu's framework: agent product-market fit emerges first in tasks that are repetitive enough to scope tightly, but complex enough that a simple automated script fails. Avoid architecture problems or novel debugging until agents mature further.
• •Model neutrality as competitive moat: Devon operates as a compound model system, dynamically routing sub-tasks across Anthropic, OpenAI, Google, and open-source models based on task complexity and cost efficiency. Wu calls this the "Switzerland" strategy — customers trust Devon to optimize price-performance rather than push a single lab's tokens. This neutrality also insulates Cognition from dependency on any single foundation model provider as the competitive landscape shifts.
• •Focus beats resources in software AI: When investors challenged Cognition against Microsoft's GitHub Copilot, Wu applied the Daniel Ek/Spotify framework — Cognition will simply care more about end-to-end software engineering than any platform company can. Startups win not by matching resources but by making concentrated bets on specific futures and executing tightly. Cognition's 75-80% enterprise revenue concentration reflects deliberate narrowing: real codebases, real teams, real output — not hobbyist demos.