Karpathy Built His Own YouTube from HN Comments with Claude Code in Two Hours

YouTube's recommendation feed is optimized for holding your attention, not for your interests. A cover with a cute cat and a headline like "Never Do THIS at the Airport" works flawlessly, and we click even though in our right mind we wouldn't watch such things. This isn't a conspiracy, just ordinary platform economics. Now there is a practical way out: write your own feed.

What Karpathy did

Andrej Karpathy, the author of the term "vibe coding," built a site called HackerTube. The idea is simple. On Hacker News, people have for years left comments with links to videos worth watching. It is a ready-made curated collection, just scattered across thousands of threads.

The prompt, as described by the author of the post, went roughly like this: read the entire Hacker News comment archive, pick out recommendations of YouTube videos, build a site from them, and base the interface on YouTube itself. The work took about two hours, and the project is already live and open. The videos on the home page really do differ from what an algorithmic feed usually serves up.

One important detail: Karpathy did nothing exotic. The author of the post says outright that Claude Code could handle such a task for any of us. The expertise here is needed not in the code but in the idea: understanding which source of signal is better than clicks.

Where the money actually went

The cost is also telling. It all fit within a regular subscription, but at API prices it would have come to roughly $200. Almost the entire sum went to labeling the Hacker News archive: the model had to go through a huge array of comments and work out which were video recommendations and which were just a mention of a link. The programming itself cost pennies.

For product managers and developers, the takeaway is simple: the cost of such a project is determined not by writing the interface but by processing the data. A clone of an interface is generated in minutes. What is expensive and slow is reading and classifying a large corpus of text.

What to do with this

First: if you need a tool for your own tastes, consider which data source could replace someone else's algorithm. It could be a comment archive, your bookmarks, newsletters, or your team chat. After that it all comes down to a prompt with a few actions: read, select, build.

Second: budget by labeling, not by code. Before running on the entire archive, it's worth running a trial sample and estimating how many tokens go into each unit of data. That way you won't be surprised by the bill if you work through the API rather than on a subscription.

Third: stay skeptical. HackerTube shows the tastes of one audience, namely Hacker News readers. For them it is a hit; for everyone else the feed may turn out boring or alien. The author of the post, incidentally, didn't end up watching anything, which makes sense: the site doesn't know his tastes because it wasn't built around them. Real personalization begins where you choose the source of the signal yourself.

And one last limitation: this is a personal experiment, not a product with a proven audience. There is no data on how people will use HackerTube. But as a demonstration it works: when creating your own video hosting site costs an evening and a couple hundred dollars in tokens, the habit of putting up with someone else's feed becomes a choice, not a necessity.