A non-programmer recreated Karpathy's HackerTube for Habr: 388 videos in one evening

YouTube recommendations are excellent at predicting what you will click next, but poor at finding the talk you will still be thinking about a year later. On October 31, Andrej Karpathy published HackerTube: 6,454 videos that Hacker News users repeatedly recommended to one another from 2007 through 2026. According to him, Claude wrote the site in Claude Code in about two hours. The author of the Habr article repeated the experiment for the Russian-speaking community and called the result "HabTube."

How it works

The data source is Habr's open archive: 302,049 articles and more than 10 million comments through February 2023. A script extracted 121,379 video links from them. The same threshold used by Karpathy was then applied: a video had to be recommended by at least three different people in three different months. A one-off spike in a hot thread therefore does not qualify, while something recommended over the years does. This produced 568 candidates.

The crucial step was having the model read the comments. Claude read each comment together with the article headline and assigned a label: strongly recommends, recommends, cites as a source, jokes, or criticizes. Without this step, a rickroll shared by 22 people across 18 different months would have topped the list. After filtering out 88 videos removed from YouTube, 84 jokes and junk entries, and 8 entries without data, 388 videos remained. For each one, the model wrote a Russian-language description and added its type, difficulty level, and an obsolescence note.

The site is a collection of static files: one index.json file and a separate comments file for each video. Likes and viewing history are stored in localStorage; there is no server-side database, and everything is hosted on Cloudflare Workers.

Six pitfalls

The list of pitfalls is more useful than the site itself because they are typical of any large-data project.

• Ilya Gusev's original dataset is gated behind acceptance of terms and returns 401. The author found an open copy and verified its schema and row count through datasets-server rather than relying on its name.
• The archive is 6.2 GB, while the laptop had 4 GB free. The solution was a rented server and reading only the required columns from Parquet in batches of 300 articles. A file containing 10,000 articles is processed in about 3 seconds.
• YouTube links come in roughly a dozen formats: youtu.be, embed, shorts, live, nocookie, and & inside HTML. A single regex solved the problem, with validation that the ID was exactly 11 characters long.
• YouTube blocks server IP addresses as bots. Metadata for 564 videos had to be collected once from a residential IP address.
• Because of the hl=ru parameter, English-language talks automatically enabled Russian subtitles. cc_load_policy=0 and loading the captions module fixed the issue.

Where it is better and worse than the original

"HabTube" has features HackerTube lacks: skipping sponsored segments through SponsorBlock—as of November 4, annotations were available for 41 videos, covering 59 segments and 42 minutes—and Russian-language descriptions even for English-language talks. Karpathy's collection has 17 times more videos, and its data is current through 2026. HabTube currently has data only through February 2023; comments from 2023–2026 are still being downloaded.

Limitations and conclusion

The descriptions were written by a model, so it may get details wrong; the original comments are attached beneath each video for verification. In Russia, a VPN is required to play the videos. RuTube and VK videos almost never met the threshold.

The practical takeaway for product teams and founders is that the recipe—an open corpus, a threshold based on independent people, an LLM as a tone classifier, and static output—can be reproduced in one evening without a team. The most valuable part is not the code but the filter: without labeling jokes and rickrolls, the result would have been a meme feed. If your community has an archive of discussions, you can build a similar collection in the same way. Just verify data availability in advance and remember that many platforms throttle server IP addresses.