cd /news/artificial-intelligence/andrej-karpathy-got-fed-up-with-yout… · home › topics › artificial-intelligence › article
[ARTICLE · art-162998] src=madrobot.blog ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

Andrej Karpathy got fed up with YouTube, so he had Claude build him a better one in two hours

Andrej Karpathy built HackerTube, a YouTube-style site that surfaces only the 6,454 videos Hacker News commenters recommended between 2007 and 2026, with Claude Opus 5.5 doing nearly all the work in about two hours. Karpathy said Claude filtered 43 million Hacker News comments, using Claude Sonnet in a Claude Code workflow to strip spam and leaving roughly 40,000 recommendation comments, which Claude then judged and summarized into cards across 150 topics. His X post about the project drew more than 4.6 million views and the blog post passed 2,800 points on Hacker News, with readers building their own versions for books, films and Apple TV.

by read5 min views3 publishedNov 3, 2026
Andrej Karpathy got fed up with YouTube, so he had Claude build him a better one in two hours
Image: Madrobot (auto-discovered)

The HackerTube logo. Image: HackerTube

Andrej Karpathy, the AI researcher who coined the term “vibe coding”, got fed up with YouTube late on a Friday night, so on Saturday morning he built his own. HackerTube shows only the 6,454 videos that Hacker News commenters have been telling each other to watch since 2007, and Karpathy says Claude Opus 5.5 did nearly all the work in about two hours.

His post about it on X has been viewed more than 4.6 million times, the blog post has passed 2,800 points on Hacker News, and readers have taken up his invitation to build their own, from versions for books and films to an Apple TV app.

A YouTube without the slop #

Karpathy says he opened YouTube at 11pm and found “an enshittified wall of slop”, with “insanely exaggerated thumbnails” and “dumb recommendations of the same tired topics”. His fix, at hackertube.org, looks like YouTube and plays videos from YouTube, but everything on it was picked by Hacker News readers.

The about page explains the rules. Every comment posted to Hacker News from 2007 to 2026 was scanned for YouTube links, and a video makes the cut only if at least three different people recommended it, with some real commentary, in at least three different months. Claude then read 39,834 of those comments and judged whether each one was recommending the video, citing it, joking about it or criticising it, before writing a card for every video: a short summary, why Hacker News likes it and whether it has aged well.

The result is sorted into 150 topics, with a home feed that mixes classics, hidden gems and picks based on what you like. There are no accounts: likes, history and Watch later live in your browser. Thumbnails are a frame from halfway through each video, not the up’s clickbait image. It is free and works in any browser, worldwide.

Its “HN canon” playlist includes Ray Dalio’s How The Economic Machine Works, recommended by more Hacker News commenters than any other video, and a half-hour film about dishwasher detergent.

“I only talk to Claude and have it do everything” #

Karpathy says Claude filtered 43 million comments from 20 years of Hacker News, running Claude Sonnet over them in a Claude Code workflow to strip out spam, leaving about 40,000 comments that recommend a video. It then wrapped the lot in a YouTube-style interface. He explained the rest when he shared it on X:

Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. Significantly better viewing experience, ~2 hours with Opus 5.5:https://hackertube.org/

  • Excellent recs without any logins (state is kept in local browser session)

  • Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads

  • Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version

  • HackerTube Longs ✊

A bit more on the process:

I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.

Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.

November 1, 2026 His only hands-on jobs were creating a Google Cloud key to fetch YouTube video details, starting a rented server and buying the domain. Claude did the rest, from deploying the site to hardening the server. The site itself warns that Claude wrote the labels and summaries, “so expect a few to be off”.

It is the latest weekend project to show how far one person can now get with an AI coding agent, after a sports fan’s 12-hour trivia game that a million people played and the vibe-coded fix that strips the ads from smart TVs.

Now everyone is building their own #

Karpathy shared HackerTube’s data as a free download on its data page, with a ready-made prompt for handing it to Claude Code or another AI agent, and told people to “point your Claude at it and ask for your own version”. They did. HackerBooks, a version for books mentioned on Hacker News, reached more than 1,800 points the next day. Wikihack did the same for Wikipedia articles, and others built HNMovies, a TikTok-style HackerTok, an open-source Apple TV app and a Fire TV and Android TV version.

One builder, richardw, who made a version for code repositories, said he usually never publishes anything: “this nudged me over the line. Thank you Karpathy!” Others saw a way out of feeds built for attention, not quality, the same problem behind the 50 million AI slop tracks Spotify pulled from its recommendations this year. One commenter, theptip, said tools like this give people “an opportunity to fully cut the cord from eyeball merchants”. The main worry in the thread was for Hacker News itself, with one commenter, po, warning that a hit like this “puts a lot of pressure on moderation of its source”.

Why it matters #

HackerTube shows that the hard part of a recommendation site is no longer the code: Claude built and hosted it in an afternoon, and the value is in 19 years of Hacker News taste. With the data free to reuse, anyone can now build a feed from the communities they trust.

Sources: Andrej Karpathy’s blog; Karpathy on X; HackerTube; HackerTube data; Hacker News; HackerBooks on Hacker News.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @andrej karpathy 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/andrej-karpathy-got-…] indexed:0 read:5min 2026-11-03 · —