HackerTube: the data work behind Karpathy’s Claude-built video site
HackerTube turns Hacker News video recommendations into a browsable catalog. Andrej Karpathy presented the project in an October 31 post, describing a roughly two-hour Claude-assisted build. The useful question is how much work went into selecting the videos, and how much of that work readers can inspect.
The distinction matters for anyone considering a similar tool. Building a familiar interface is one task. Deciding which links deserve a place in it is another, with different costs and failure modes.
How HackerTube chooses videos
The project’s About page lists 6,454 videos drawn from 39,834 comments. Its inclusion rule requires links from at least three people across three months. Claude labels whether comments recommend, cite, joke about or criticize a video. Ranking then favors recommendations spread across separate months. These are the project’s stated methods, not an independent quality audit.
During a November 2 browser check, a search for distributed systems returned relevant results. A Martin Kleppmann lecture page displayed seven comment excerpts, stance labels and links back to Hacker News. We inspected those pages, without testing video playback or recommendation quality.
What the data export includes
| File | Contents described by About |
|---|---|
| videos.jsonl | Video records, summaries and topics |
| mentions.jsonl | Comment text, source links and stance labels |
| DATA.md | Field definitions, examples and licensing guidance |
About offers a 17.8 MB data archive under ODC-By 1.0, with credit to HackerTube. It excludes YouTube descriptions, view counts and durations. This is a data export; we have not verified a public application-code repository.
What the reported build effort includes
In a Hacker News reply, Karpathy reported Claude’s tally of 504 million input tokens, around 90% cached, and 7.6 million output tokens. He said more than 90% of usage went to Sonnet comment labeling, while the app and deployment used roughly one million Opus tokens. He used a subscription; those figures do not establish a dollar cost.
His blog also identifies manual setup: a YouTube metadata API key, a virtual server and a domain. The reported build time should not become a promise that another developer can reproduce the result on the same budget or schedule.
What to borrow from the workflow
Start with a bounded collection and retain the evidence behind each selection. For a reading list or lecture finder, distinguish praise from criticism before treating a mention as a recommendation. Check a sample of labels against the original comments, and publish the collection’s refresh date.
Hacker News preferences will not represent every audience. Source links make those preferences easier to examine, but do not guarantee that a video is accurate, current or available. A maintained discovery tool needs those checks after its first deployment, too.
AI-generated editorial illustration by ByteForward. HackerTube wordmark used for identification; no endorsement implied. Not a product screenshot. Identity source: HackerTube.
