“Engineers are becoming sorcerers” | The future of software development with OpenAI’s Sherwin Wu

What It Covers

Sherwin Wu, head of engineering for OpenAI's API and developer platform, reveals how 95% of OpenAI engineers use Codex daily with AI writing nearly all code. He discusses the transformation of software engineering into agent management, the one-person billion-dollar startup future, why listening to customers can mislead AI product development, and untapped opportunities in business process automation beyond Silicon Valley's focus.

Key Questions Answered

• •AI code generation at scale: At OpenAI, 95% of engineers use Codex daily, with AI writing nearly all initial code that engineers then review. Engineers using Codex open 70% more pull requests than those who don't, and this productivity gap continues widening. The company maintains a team running a 100% Codex-generated codebase as an experiment, discovering that most agent failures stem from insufficient context documentation rather than model limitations, requiring better tribal knowledge encoding.
• •Engineering role transformation: Software engineers transition from writing code to managing 10-20 parallel agent threads simultaneously, resembling tech leads overseeing teams. The job shifts toward providing clear specifications and steering AI agents rather than manual coding. This mirrors the "wizard" metaphor from the 1980 programming textbook SICP, where engineers cast incantations that execute tasks, with current reality approaching the Sorcerer's Apprentice scenario requiring skilled oversight to prevent runaway automation.
• •Manager leverage expansion: Engineering managers can oversee significantly larger teams than the traditional six-to-eight person limit by using AI tools for organizational knowledge synthesis. ChatGPT connected to GitHub, Notion, and Google Docs enables rapid performance review research and team status understanding. The management philosophy shifts toward spending over 50% of time with top 10% performers who maximize AI tool leverage, as these individuals become exponentially more productive and establish best practices for entire organizations.
• •Customer feedback paradox: In rapidly evolving AI fields, blindly following customer requests leads to local maxima solutions that become obsolete. Models improve so quickly they "eat your scaffolding for breakfast"—vector stores and agent frameworks that seemed essential in 2023 became less relevant as models gained native capabilities. Product builders must design for where models will be in 12-18 months, not current capabilities, accepting that 80% functionality today may reach full capability with next model releases.
• •Billion-dollar startup ecosystem: The one-person billion-dollar startup concept triggers second and third-order effects: dramatically lower software creation barriers spawn hundreds of thousands of smaller startups building vertical-specific tools. This creates a golden age of B2B SaaS where $10-50 million businesses become common, excellent for founders but challenging for venture capital seeking 100x returns. The ecosystem shifts toward platforms supporting micro-companies rather than traditional venture-scale consolidation.
• •Business process automation opportunity: Silicon Valley underestimates the massive market for automating repeatable business processes outside open-ended knowledge work. Unlike software engineering's creative tasks, most global work follows standard operating procedures in support, operations, and enterprise functions. These deterministic, repeatable workflows integrated with business systems represent untapped AI application territory larger than the engineering productivity space that dominates current discourse and investment focus.