A Friday-night frustration
Andrej Karpathy wrote on his personal blog on Saturday, October 31, that at 11 p.m. the night before he had opened YouTube looking for something to watch and found exaggerated thumbnails and the same tired recommendations. He allowed that the algorithm probably works for the median person, but said it does not work for him. So the next morning he gave Opus 5.5 a task: build our own YouTube from the videos people have recommended on Hacker News.
The result is HackerTube.
40,000 comments, 6,454 videos
As Karpathy describes it, Opus 5.5 handled the whole pipeline: it downloaded all Hacker News comments and filtered them, then ran Sonnet over them in a workflow to remove spam and low-grade content, leaving about 40,000 comments that recommend a YouTube video. Altogether that comes to 6,454 videos recommended by a few HN users between 2007 and 2026, across 150 topics. Opus then wrapped them in a YouTube-style interface, got a domain, deployed to a VPS, and secured, hardened and optimized it. He says a project like this "takes ~2 hours now."
A few design choices deliberately differ from YouTube:
- Thumbnails use the frame at the 50% point of the video, not the creator's uploaded thumbnail;
- The For You feed needs no account, keeping preference state only in the local browser session;
- Keyboard shortcuts are supported;
- An export of the site's data is available; Karpathy suggests handing it to your own agent to make your own version.
Why it is worth a look
The interesting part is not that there is one more video site, but an increasingly common pattern: a strong model writes the code, a cheaper model cleans data in bulk, a personal idea that once took weeks fits into one morning, and the data is published so others can reproduce it. The post quickly reached the Hacker News front page and, when this article was verified, had more than 700 points and over 200 comments. Bear in mind that coverage reflects the HN community's tastes, which lean toward technology and science, and that filtering quality depends on the models' judgments; the site does not publish its exact selection criteria.
via: Karpathy's blog: HackerTube, HackerTube, Hacker News discussion