Scott Wu - Building Cognition - [Invest Like the Best, EP.402]

Episode Summary

Executive Summary: Patrick O’Shaughnessy interviews Scott Wu, CEO of Cognition, about Devin, an AI software engineer that automates end-to-end coding work. Wu argues software is bottlenecked by labor, and agentic AI will expand what gets built, who can build it, and how fast teams ship.

Main Topics: What Cognition and Devin do: Devin is an autonomous software engineer that handles bugs, PRs, testing, and tool use. From copilots to agents: Wu contrasts synchronous autocomplete tools with asynchronous systems that complete full tasks. Software engineering history: He frames programming as the evolving art of telling computers what to do, from punch cards to AI. The future of software creation: Wu expects natural language to increasingly replace coding for many tasks over 5-10 years. Building with foundation models: Cognition focuses on product, workflow, and tooling around models rather than raw model IQ alone. Business model and pricing: Devin is priced by usage to reflect compute and value delivered on end-to-end tasks. Startup lessons and team culture: Wu emphasizes speed, customer closeness, hiring quality, and the advantage of trusted cofounders.

Key Arguments: Devin automates full software workflows: reproduce bugs, inspect logs, test fixes, and open PRs. Wu says Devin is a junior engineer today, not a replacement for senior architects. Agentic tools can create ~10x leverage on tasks they can fully own, versus 10-20% from copilots. Software has long been supply-constrained, so cheaper engineering should expand total software demand. Cognition wins by integrating with real tools and messy workflows, not solving code in a vacuum. Model progress alone is insufficient; product design, memory, and workflow integration matter. Pricing should follow usage and value because agent tasks consume far more inference than Q&A. Wu thinks AI will democratize software creation by letting non-programmers specify what they want.

Data Points: Company age: just around one year ago - Cognition was founded about a year before the interview. Devin pricing: $500 bucks a month - Default plan for engineering teams. Productivity gain: eight to 12 hours - Internal studies found 1 hour with Devin equaled 8-12 hours without it. Copilot-style improvement: 10 to 20% - Wu estimates code assistants improve engineers by this amount. Devin task leverage: 10x - For tasks Devin can handle, Wu says it is more like a 10x boost. Competitive programming target: one to two years - Wu predicts AI will beat the world's best competitive programmer in this window. Foundation model query cost: one millionth of a dollar - Example cost for a single GPT-style Q&A query at API pricing. Usage unit: AC use agent compute units - Cognition bills by agent compute usage. Devin hourly cost: 8 to 12 bucks - Approximate cost per hour of Devin work. Team size: about 20 people - Wu describes Cognition's team size during the interview. Founders on team: 14 of them - Of the 20-person team, 14 are former founders. AI progress timeline: five to ten years - Wu's estimate for natural-language software creation for many users. LLM milestone: ChatGPT in November of 2022 - He cites this as the moment the world realized what was possible. Launch timing: January, February, March, April, May - Wu describes Lunch Club's rapid early cadence.

Pivotal Quotes: "software engineering is just very messy" — Scott Wu: He explains why building Devin requires workflow and infrastructure, not just model capability. "we've been playing in Minecraft survival mode for so long and now it's going to be time for Minecraft creative mode" — Scott Wu: His metaphor for AI moving humans from execution constraints toward pure creation. "The biggest distinction is actually not whether we are number one in the world or only top 0.1% human or something like that" — Scott Wu: He argues practical impact matters more than benchmark supremacy.

Implications: The open question is adoption speed: teams should prepare for agent-ready workflows now, because product and distribution may matter as much as model intelligence.