Y Combinator's CEO says he ships 37,000 lines of AI code a day. An independent developer audited the result.

When Y Combinator CEO Garry Tan posted on X that he was shipping 37,000 lines of AI-assisted code a day across five projects, a developer went looking at the actual output. The site Tan maintains as a public calling card, garryslist.org, delivers a newsletter and a blog. According to an independent audit published on X by a developer posting as Gregorein, loading the homepage pulls 6.42 megabytes across 169 requests and serves roughly 78,400 lines of production code.

The audit's numbers describe what happens when you optimize for LoC, a productivity metric the engineering community had been contesting for decades before AI assistants arrived. A cold load of garryslist.org pulls a 6.42-megabyte payload across 169 requests for a site whose job is to push out a newsletter and host short posts. The auditor characterized the output as "AI slop"; that label is Gregorein's, not the article's. The underlying numbers are reproducible and stand on their own.

Tan framed his streak as "an insane week for agentic engineering," with a 72-day shipping streak across five projects. An earlier Forbes profile from April 2026 had already documented the workflow under a softer banner, calling Tan "the YC chief who codes 10,000 lines a day" and attributing the volume to a "simple secret." The 37,000-line claim is the escalation. The Gregorein audit is the first time someone other than Tan has tried to count what comes out the other end.

Gregorein ran a network-panel inspection of garryslist.org and tallied the asset weight and request count for a cold homepage load. For comparison, a comparable personal site built with conventional static-site tooling often loads in well under a megabyte from a single document request. The audit is a single developer's measurement of one site, not a benchmark, but the methodology is in the public record.

The Hacker News thread that surfaced the audit turned quickly to the harder question: what does LoC actually measure when an AI assistant is doing most of the typing? Commenters noted that AI scaffolding tends to generate wrapper classes, defensive null-checks, and verbose configuration blocks that all count toward LoC without adding product surface area. Others pointed out that "good enough for now, ship and debug later" is invisible to readers on a MacBook with fiber internet and a fast phone. The cost shows up on cold devices, slow networks, and battery-constrained laptops, exactly the conditions startup founders rarely test in.

The stakes live in Tan's role. Y Combinator is the most influential startup accelerator in the United States, and the norms Tan models, shipping speed, output volume, AI-first scaffolding, travel downstream into the roughly five hundred companies it funds each year and, via those companies, into the broader founder community. When the headline metric is "lines of AI code shipped per day," the floor for "done" rises to whatever the metric rewards. The artifact's quality under developer scrutiny becomes invisible to the people setting the example.

Tan has not, as of this writing, responded to the audit's specific findings in public. A charitable reading is that the bloat reflects experimentation rather than a finished product. A less charitable reading is that the audit is the natural consequence of optimizing for a metric that was already broken before AI assistants arrived to inflate it. Fast Company covered the broader agentic-engineering moment around Tan, and an earlier Hacker News discussion of his workflow drew similar critiques of LoC as a productivity signal.

The audit pins a number to the gap between "shipping" and "having built something that holds up." The line count keeps climbing. The next independent audit of a Tan project is the test of whether the artifacts keep pace.