Viral Social Media Roast Calls Out YC CEO Garry Tan’s AI-Generated Code Bloat

Viral social media roasts of public tech figures are always a draw, and Y Combinator CEO Garry Tan became the target of a particularly notable one this Wednesday.

Tan, a former engineering manager at Palantir who has emerged as a prominent vocal advocate for rapid AI development, recently went public with impressive-sounding metrics for his work with AI coding agents. In a Monday X post, he shared that he and his AI tools deploy 37,000 lines of code per day across five separate projects. “Absolutely insane week for agentic engineering,” Tan wrote, adding in a follow-up that he had kept up a 72-day streak of consistent code deployments.

Two days later, a senior software engineer and Polish game developer who goes by the username Gregorein decided to audit the real-world output of all that code shipping. After pulling the public code from Tan’s AI-focused personal blog, Gregorein shared his findings in a viral response X post: “Here’s what 78,400 lines of AI slop code actually looks like in production.”

Gregorein, who holds a master’s degree in computer science, a professional Polish engineering qualification, and has 13 years of industry experience, confirmed his observations by reviewing all downloaded public site files in a single session with Anthropic’s Claude AI. He uncovered a long list of unnecessary bloat and basic inefficiencies, including:

• When a visitor loads the site, their browser sends 169 separate server requests for different assets, totaling 6.42 megabytes of data transferred. For context, Y Combinator’s own minimalist Hacker News homepage only sends 7 requests, totaling just 12 kilobytes.

• 28 full developer test files (code used by engineers to validate work during development) are pushed directly to production and downloaded by every visitor, adding 300 kilobytes of unused development scaffolding that end users never need.

• The site pre-loads 78 separate JavaScript controllers for off-page features including AI image generation, voice extraction, and video tools — none of which appear on the site’s homepage. All 78 files are downloaded regardless, “just in case” a user navigates to those pages later.

• The site’s bear logo illustration is served in eight different file formats, including one completely empty 0-byte file that made it all the way to the live production site.

• Most images are served as large, uncompressed legacy PNG files (some nearly 2 megabytes apiece), even though modern browsers explicitly request smaller, more efficient modern file formats. Just two of these uncompressed images waste roughly 4 megabytes of bandwidth, compared to 300 kilobytes total if they were converted to a modern format.

• Additional issues include duplicate page content, an entirely empty Cascading Style Sheets (CSS) file, a full large rich-text editor loaded on a static read-only page, missing accessibility descriptions for images, and analytics code intentionally routed through a proxy to bypass users’ ad blockers — with an in-code comment explicitly confirming this work-around.

Gregorein emphasized that his review only covered public front-end code viewable in visitors’ browsers, and did not include an audit of back-end or database code. Even without looking at non-public code, the public-facing user experience makes clear that the site is riddled with obvious bloat, wasted resources, and basic rookie mistakes, he concluded.

The exchange highlights a core ongoing debate about AI-powered development: while AI coding tools make it extremely easy to churn out huge volumes of code quickly, code quality is still what matters most. Put simply, quantity does not equal quality. Non-technical users can now use plain language prompts to build websites, apps, and new features in a fraction of the time it once took, but if that unvetted AI code is pushed directly to production, it can cause obvious performance failures, open up dangerous security vulnerabilities, or create hidden technical debt that requires engineers to spend far more time fixing problems later.

In a direct message to Fast Company Thursday, Gregorein explained that when individual deployments add massive amounts of new code at once, quality tends to degrade exponentially. Each new round of AI-generated changes requires more and more rework just to keep the build stable, he added. “Right now we’re in a moment where AI lets you generate code faster than any human can review it, and the answer from people like Garry seems to be ‘so stop reviewing,’” Gregorein said.