A Distinguished Engineer Stopped Reading Code — AI Assistance Is Rewiring What Seniority Means
Philip Su, who reached Distinguished Engineer IC9 at Meta before joining OpenAI, argues in a new interview that code reviews are becoming liabilities rather than best practices in an AI-assisted development world. The conversation explores whether the traditional individual contributor career path still makes sense when AI systems can outpace human code comprehension.
Original sourceThe conversation around AI and software engineering has mostly focused on junior developers — whether they're being displaced, whether they're still learning effectively, whether the entry-level pipeline is broken. A new interview challenges that framing by starting at the top of the IC ladder instead.
Philip Su, who reached Distinguished Engineer (IC9) at Meta and has since worked at Microsoft and OpenAI, argues that the transformation is happening at every level — and that senior engineers are not immune to having their core practices disrupted. His central claim: that code reviews, long considered a best practice and a cornerstone of engineering quality culture, are becoming operational liabilities as AI-generated code volumes exceed what humans can meaningfully audit.
The argument is not that code quality doesn't matter, but that the review process itself — human engineers reading and commenting on diffs — was designed for a world where humans wrote the code at human speed. AI-assisted codebases can generate tens of thousands of lines per day, and the cognitive overhead of traditional review practices doesn't scale. Su's own response has been to stop reading code and instead focus on higher-order concerns: system properties, interface contracts, and observability signals.
The interview also explores what Su calls "lights-out codebases" — systems where AI agents write, test, and deploy code with minimal human involvement in the individual-line decisions. He argues this model is closer than most engineers want to admit, and that the IC career path (which rewards deep individual expertise in reading, writing, and reviewing code) needs to be rethought in light of it.
The piece generated significant HN discussion, with engineers divided between those who see this as overreach (reading code remains the only way to understand what a system actually does) and those who've already empirically reduced review time in favor of property-based testing and integration signal monitoring. The tension isn't resolved, but the conversation is overdue.
Panel Takes
The Builder
Developer Perspective
“I find myself reading less diff and checking more behavior and test coverage than I used to. Su's framing isn't fully right, but it's pointing at something real: the cost-benefit ratio of line-by-line review is shifting as AI-generated code becomes a larger fraction of what ships.”
The Skeptic
Reality Check
“This is a provocative thesis from someone at a company with extraordinary resources and safety nets. Most engineering teams don't have the observability infrastructure or testing culture to stop reading code safely. The advice generalizes poorly and risks giving cover to the teams that needed code review most to abandon it.”
The Futurist
Big Picture
“The IC career ladder was built around human bottlenecks in code production. Once that bottleneck lifts, the entire structure of how we reward, evaluate, and promote engineers needs rethinking. Su is one of the few people senior enough to say this publicly — expect this conversation to get louder in 2026.”