# Karpathy built a whole YouTube clone in two hours and barely wrote any code

**URL:** <https://forum.kirupa.com/t/karpathy-built-a-whole-youtube-clone-in-two-hours-and-barely-wrote-any-code/683911>\
**Category:** talk\
**Created:** [November 2, 2026, 5:14pm UTC](https://forum.kirupa.com/t/karpathy-built-a-whole-youtube-clone-in-two-hours-and-barely-wrote-any-code/683911 "2026-11-02T17:14:10Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![PipMuse](https://yyz1.discourse-cdn.com/flex011/user_avatar/forum.kirupa.com/pipmuse/32/34326_2.png) [@PipMuse](https://forum.kirupa.com/u/PipMuse)\
**Post date:** [November 2, 2026, 5:14pm UTC](https://forum.kirupa.com/t/karpathy-built-a-whole-youtube-clone-in-two-hours-and-barely-wrote-any-code/683911/1 "2026-11-02T17:14:10Z")

</div>

Okay, this one got me. Andrej Karpathy got fed up with his YouTube homepage one Friday night — he called it a wall of slop — and by Saturday morning he’d basically built himself a replacement.

It’s called HackerTube, and the idea is dead simple: instead of an algorithm picking what you watch, every video on it got there because someone recommended it in a Hacker News comment. Claude Opus 5.5 did nearly all of the work in about two hours. Karpathy says his own contribution was three chores: a YouTube API key, a server, and buying the domain.

The numbers are kind of wild. Claude sifted through Hacker News comments from 2007 to 2026, narrowed it down to around 40,000 comments that actually recommend a video, and ended up with 6,454 videos sorted into about 150 topics. It even skips the clickbait thumbnails and shows a frame from the middle of each video instead. No account needed, and you can export the whole dataset and build your own version.

Fair caveat: the two hours and the token counts are Karpathy’s own account, not something anyone’s independently reproduced yet. Still, it topped Hacker News over the weekend, and people are already cloning the idea — HackerBooks, Wikihack, that sort of thing.

Here’s the writeup I read: [Karpathy builds HackerTube in 2h with Claude Code](https://www.thenewway.ai/karpathy-builds-hackertube-in-2h-with-claude-code/)  
And the site itself: [https://hackertube.org](https://hackertube.org)

So here’s my question: would you actually trust 6,454 picks from forum strangers over YouTube’s algorithm? And if you had two hours and an agent that good, what would you build first?