37,000 lines of code per day. That's what Y Combinator CEO Garry Tan says he's shipping with the help of AI coding tools.
To put that number in perspective: a senior software engineer writing production code by hand typically produces 50 to 150 lines of meaningful code per day. The entire Linux kernel, built over decades by thousands of contributors, is about 28 million lines. At Tan's claimed pace, he'd match that in roughly two years, solo.
The Number Doesn't Mean What You Think
The gap between "lines generated" and "lines that matter" is enormous. Anyone who's used Cursor, Claude Code, or GitHub Copilot knows the pattern: the AI spits out hundreds of lines in seconds, but you spend the next hour reviewing, testing, and rewriting half of it. Raw line count is the worst possible metric for code quality or productivity.
Tan's number almost certainly includes boilerplate, generated tests, scaffolding, and auto-completed patterns. That's not a criticism of AI coding tools - they're genuinely useful for exactly those tasks. But calling it "shipping 37,000 lines" conflates output volume with engineering value.
What It Actually Tells Us
The interesting part isn't the number itself. It's that the CEO of the most influential startup accelerator in tech is spending his days deep in AI-assisted coding rather than just evangelizing it. Tan is clearly building real projects with these tools, and his willingness to put a specific (if inflated) number on it signals how normalized AI coding has become at the top of the startup world.
For the thousands of YC founders watching their CEO, the message is clear: if you're not building with AI coding assistants, you're already behind. That pressure will trickle down to every startup pitch deck and hiring decision in the YC orbit.
The Metric We Actually Need
The AI coding community desperately needs better productivity metrics than lines of code. What percentage of AI-generated code survives code review unchanged? How many bugs per AI-generated function versus human-written ones? How much time does review and debugging add back?
Until those numbers exist, claims like "37,000 lines per day" are marketing, not measurement. The tools are real and genuinely useful - but we should talk about them honestly instead of reaching for the most impressive-sounding stat.