Philip Su has held senior engineering roles at Microsoft, Meta (where he reached Distinguished Engineer at IC9), and OpenAI. For the last four months, he hasn't written a single line of code himself. Instead, he spends 40 hours a week orchestrating AI agents, primarily Claude Code CLI, to build Superphonic, his AI-powered podcast player.
His thesis is blunt: the individual contributor role is over. Engineers who thrive will be the ones who stop writing code and start managing the systems that write it. Su frames the shift through an offshoring analogy. Early outsourcing gave you worse code but cheaper. AI is the same trade today, except the quality ceiling is rising fast.
Su introduces what he calls "lights-out codebases," borrowing from data centers that run with zero human workers. Codebases where no human ever looks at or touches the source. He draws a parallel to chess engines. Stockfish, free on your phone, beats the world's best players. Having a human grandmaster review every move would be absurd. We'll reach the same point with code, he argues. Human intervention becomes a risk, not a safeguard. If developers produce 5x to 20x more code with AI assistance, traditional code review becomes impossible.
This is already happening. Cognition's Devin handles end-to-end coding tasks autonomously. Replit's agents generate complete applications from natural language. Microsoft's Power Platform and Salesforce's Einstein Copilot run AI-generated business logic in production with minimal human inspection.
Simon Willison pointed out in the Hacker News discussion that developers already work with code they haven't personally reviewed. Trusting AI output isn't so different from trusting another team's code. Su has built trust with specific Claude model versions for common code categories. Same way you build trust with a colleague over time.
His advice to engineers: stop reading about AI and start building with it. "Trust no one, do the work yourself," he says. The judgment needed to evaluate AI output only comes from hands-on experience. He thinks most people are wrong about timing, expecting changes to take years when improvements are compounding faster than expected. His own career backs up the urgency. He went from distinguished engineer to moving six tons of packages a day at an Amazon warehouse during Peak 2021. That experience taught him never to laugh when a robot fails. Because V2 is coming.