Karpathy Had Claude Read 43 Million HN Comments and Filter Out 6,454 Videos Worth Watching

Andrej Karpathy built a website called HackerTube over the weekend.

The motivation was straightforward: he was annoyed by YouTube's recommendations. So he had Claude filter the 43 million comments Hacker News has accumulated over 20 years, ultimately leaving 6,454 videos that had been repeatedly recommended. According to him, the viewing experience is noticeably better. The whole process took about two hours with Opus 5.5.

HackerTube URL: https://hackertube.org/

The site makes a few clear choices:

• No login; state is kept in the local browser session.
• Thumbnails are representative frames from the middle of the video, not covers carefully A/B tested by uploaders.
• The comment-analysis data is packaged as a zip file that can be handed to Claude so it can generate its own version.
• The navigation includes a Longs entry for finding long videos.

Karpathy said he used the Mac desktop version of Claude Code by default, with Opus 5.5 and Medium effort. For the entire project, he only talked to Claude and had it do everything. He personally did only three things: create an API key in the Google Cloud UI to query YouTube video metadata; spin up a VPS; and buy a domain. Once the VPS was up, he gave the IP to Claude and had it update, harden, deploy, and optimize the code. Claude did it all.

In the replies, someone seized on this point: writing code is not unusual; what is outrageous is that Claude also did the deployment. Karpathy put it more directly himself: many projects that used to take weeks or even months can now be few-shotted for breakfast. He also asked: what else have we not thought of?

Dan McAteer posted the first video recommended by the algorithm: an episode of the Lex Fridman podcast discussing UFOs, aliens, fighter jets, and aerospace engineering, with 6.9 million views. He said this was exactly the kind of content he wanted to watch.

One commenter summarized it more sharply: Hacker News has been arguing for 20 years, and along the way it grew a recommendation algorithm that suits people's tastes better than YouTube's. Someone else said the agent that read all 43 million HN comments deserves sympathy, because it was equivalent to swallowing decades of “well, actually.”

Pranav P R likes the site's filter options, since filtering by content type is faster. The screenshot shows different filtering dimensions.

Someone else in the comments responded with a keyboard-typing video; the vibe was basically, another weekend project has shipped.

Some people also reported that the homepage / would not load and that they had to go to /browse. Max Hager said search was a bit slow and that he even found Karpathy's own videos in there. These are minor issues, but they reveal something: the problem with recommender systems is not necessarily a lack of data; more often, the objective is drowned out by click-through rate, dwell time, and ads. HN comments are not a clean dataset; they contain arguments, bias, repetition, and also judgments accumulated over twenty years. Claude did not invent a new recommendation algorithm; it surfaced videos that a bunch of people had repeatedly voted for with their feet.

Someone has already wished for the next project: use the same method to clean up Amazon, removing sponsored placements, ads, and subscribe and save. This direction is not new, but HackerTube demonstrated the path: first find a corpus with long-term human discussion, have a model filter it, and then turn the results into a small tool that requires no login.