Andrej Karpathy got tired of YouTube's recommendations, so he pointed Claude at 20 years of Hacker News comments and asked for something better. Two hours later he had a working site with 6,454 videos on it. Most of his time went on the parts a model can't handle: accounts, a server, a domain.
What HackerTube Recommends, and What It Refuses To
HackerTube, live at hackertube.org, does one thing. It shows YouTube videos that Hacker News users have recommended more than once, across different years.
That constraint is the whole product. YouTube's own recommendations chase watch time, which tends to surface whatever holds attention for thirty seconds. Hacker News comments are a different signal: someone took the trouble to paste a link into a thread and say why it was worth watching.
Karpathy announced the project on X on November 1. The post passed 3.2 million views within 15 hours, which is what happens when a well-known AI researcher ships something on a weekend.
The ranking rule is unusual. Videos are ordered by how many separate months they were recommended in, not by total mentions or upvotes. A video that got one enthusiastic thread in 2014 and nothing since ranks below one that keeps coming back.
Five Filter Steps Turn 100,000 Links Into 6,454 Videos
The site's about page lays out the pipeline, and the numbers explain why it took an agent rather than a weekend of manual work.
Step one: scan every Hacker News comment from 2007 to 2026 for YouTube links. That produced more than 100,000 distinct videos.
Step two: keep only videos that at least three different people recommended, in at least three different months, with an actual comment attached. That cut the pool to roughly 7,000.
Step three is where the judgment happens. Claude read every comment alongside the title of the story it appeared under, then classified the stance: recommendation, citation, joke, criticism. It labeled 39,834 comments this way. A link posted as the punchline of a joke shouldn't count as an endorsement, and a link dropped to mock a video shouldn't count either.
Step four: Claude wrote a card for each surviving video with a summary, a note on why Hacker News liked it, a genre, and an assessment of whether it holds up.
Step five: 6,454 cards covering about 150 topics.
The filtering is the part worth studying, because most link aggregators treat a link as a vote. This one tries to distinguish "you should watch this" from "look at this nonsense," which is a language task rather than a counting task.
Who Did What: One Human, One Opus, a Team of Sonnets
Karpathy's account of the division of labor is specific, and it's the most useful part of the story for anyone trying to repeat it.
He did three things, all of them involving a browser and a payment method. He applied for a Google Cloud API key to pull YouTube metadata, rented a server, and bought the domain.
Claude Opus 5.5 did the rest, working in the desktop version of Claude Code on a Mac at medium effort. That included the data pipeline, the search and recommendation logic, the site itself, and the deployment. Karpathy handed over the server's IP address and told the model to update, harden, deploy and tune.
The comment labeling, the video cards and the topic list were split across several Claude Sonnet 5.5 agents, coordinated by Opus. That's the pattern people mean when they say agent orchestration: one model plans and delegates, smaller or cheaper models grind through the repetitive reads, and the planner assembles the result.
It's also a division of labor that maps to cost. Reading 39,834 comments is a lot of tokens. Reading them with a frontier model would have been wasteful when a smaller model can classify a comment's stance just as well.
The HackerTube Dataset Ships With Instructions for Agents
The data ships as a 17.8 MB zip, ODC-By 1.0, exported October 31, 2026.
Inside are four files: videos.jsonl with 6,454 rows, mentions.jsonl with 39,834 rows, related.jsonl with 3,778 rows, and a taxonomy of 150 topics. The mentions file is the big one at 34.7 MB uncompressed, because it holds the comment text.
One thing is missing on purpose. YouTube's API terms don't allow redistributing view counts, durations and similar metadata, so the archive carries video IDs only. Fetching them takes a free API key and about 130 requests.
The part that makes the archive unusual is DATA.md, a file written for AI agents rather than humans. It explains how the files relate, where the traps are, what a sample record looks like, and how to walk a user through retrieving the YouTube metadata.
Karpathy's post ends with an instruction: point your Claude at it. The dataset is designed to be handed to an agent as a starting point for another project.
What HackerTube's Two Hours Leave Out
Does that mean anyone can ship a site in an evening? The two-hour figure is Karpathy's own account, and nobody has reproduced it. Treat it as a report from the person who did the work, not a benchmark.
What it does show is where the time went. Not typing code. All three of the human tasks were procurement: an API key, a server, a domain. The model handled everything that could be expressed as a sequence of file edits and commands.
That's a real shift in what one person can build in an evening, and it's narrower than the headlines suggest. The result depends on a specific model, a specific coding environment, and a dataset that already existed. Hacker News has been accumulating comments for two decades. There's no equivalent archive for most ideas.
The other limit is legal and practical. Hacker News comments and YouTube metadata each come with terms of use. A personal weekend project can sit comfortably inside them. A commercial clone would need to look harder at both sets of terms, and probably at a lawyer's invoice too.
And the ranking is an editorial judgment, not a quality score. It measures what a particular community kept recommending, which is a specific kind of taste and not a general quality score. The site is honest about that, since the rule is printed on the about page.
Why the HackerTube Build Is a Useful Example Anyway
Strip away the personality and the project is a clean demo of a pattern that's becoming common. A person defines a filter, an agent does the reading, a second layer of agents does the volume work, and the human handles accounts and infrastructure.
The valuable skill on display isn't prompt writing. It's deciding what counts as a recommendation, then expressing that rule precisely enough that 39,834 comments can be sorted against it.
HackerTube is worth a visit for the videos. For anyone building with agents, the about page and the DATA.md file are the more interesting read.
