YouTube recommendations are too annoying, so he used Claude to surface 6,454 videos from 20 years of HN comments

In a tweet retweeted by Rohit Ghumare, Andrej Karpathy built a website called HackerTube over the weekend. The motivation was straightforward: he found YouTube recommendations annoying. The approach was just as direct: he had Claude filter 43 million Hacker News comments from the past 20 years and keep 6,454 videos that were repeatedly recommended in discussions. The site is https://hackertube.org/.

Opening the screenshot shows a left sidebar with Home, Longs, Liked, History, Watch later, Categories, Playlists, and Settings. On the right is a video list; thumbnails use a representative frame from the middle of the video rather than the uploader's polished A/B-tested image. Karpathy says this viewing experience is noticeably better. The site requires no login, and state is stored in the browser's local session. The data is also packaged as a zip; give it to Claude and it can build your own version along the same lines. There is also HackerTube Longs, specifically for long videos.

The process is more worth seeing than the result. Karpathy used the Mac desktop version of Claude Code, with Opus 5.5 by default and Medium effort. He basically only talked to Claude and let Claude do the work. Only three things were manual: creating an API key from the Google Cloud UI to query YouTube video metadata, spinning up a VPS for the service, and buying a domain. Once the machine was up, he gave the IP to Claude and had Claude update, harden, deploy, and optimize the code. According to him, Claude did it all. The whole project took about two hours.

The lesson isn't the old line that AI can write code. Karpathy's filtering source was videos that repeatedly appeared in twenty years of HN comments, not YouTube click-through rates. Recommendation systems turn thumbnails into A/B test artifacts; HackerTube uses a middle frame. One optimizes for short-term clicks, the other filters by long-term discussion. The effect may not hold for everyone, but at least it shows that recommendations can use a different signal. Many projects that used to take weeks or months can now be solved with a few-shot approach. Karpathy's point, roughly, was: what else have we not thought of?

HackerTube's data and code approach are both out in the open. Anyone who wants to try it can first look at the recommendations on the site, then hand the zip to Claude and modify a version. Note that you still need to prepare the API key, VPS, and domain yourself. Don't treat it as a universal recommender; it just brings HN's long-term taste into a video page. For developers, the value of this weekend project is the workflow: conversation, deployment, hardening, optimization—all connected. For ordinary users, at least they can see a few fewer thumbnails pushed by the algorithm.