Karpathy built HackerTube over a weekend: 6,454 videos from 43 million HN comments

A familiar situation: you open YouTube, and the feed serves up whatever holds your attention rather than what is worth watching. Andrej Karpathy approached the problem like an engineer and built HackerTube over a weekend. It is a collection of videos that the Hacker News community has recommended over two decades.

What exactly was built

According to Karpathy, Claude filtered 43 million HN comments spanning 20 years down to 6,454 videos that people there “keep recommending.” The entire project took about two hours of work with Opus 5.5. He describes the result as a significantly better viewing experience than YouTube recommendations.

The post received more than 3.5 million views. That is not just because of the author's name. The idea is immediately understandable, and almost everyone who watches technical videos is tempted to replicate it.

How it was probably built

Karpathy did not disclose implementation details in the post, so what follows is our reconstruction, not a description of his code. The logic is straightforward: people constantly leave links to videos in HN comments, accompanied by phrases such as “a must-watch.” If the same link resurfaces repeatedly, across different topics and over many years, that is a strong signal. It cannot be manufactured with a clickbait thumbnail, because the recommendation comes from someone who has already watched the video and considered it worth mentioning.

Here is a simple example. A lecture on compiler design might be mentioned in programming-language threads in 2012, 2016, and 2023. That tells the YouTube algorithm nothing. For a comment-based filter, it is an obvious candidate for the list.

This work cannot be done manually across 43 million comments. For an LLM, however, it is a typical task: read a large body of text, extract mentions, remove noise, and judge where a recommendation is genuine and where a link appeared in a thread by chance. This is exactly the kind of format that suits agentic coding: what is needed is a small, disposable tool, not a product built to last for years.

Why this matters to developers and product teams

First, it shows how the cost of “hacker” projects has changed. According to the author, a task that previously required a week of scripting, parsing, and manual cleanup can now fit into two hours and one evening. The barrier to building personal tools has fallen.

Second, it is a signal for teams building recommendation systems. Community curation once again looks like a viable alternative to behavior-based personalization. People are tired of feeds optimized for time spent in the app and are willing to watch material vetted by other people.

Third, it provides a template for your own experiments. Take an archive in which people have left evaluations for years: comments, discussions, tickets, or reviews. Ask a model to extract recurring recommendations and turn them into a showcase. This is a realistic use case for an internal team knowledge base or an educational catalog within a product.

Limitations

It is important to be clear about what we do not know. There is no quality metric: “significantly better” is the author's subjective assessment. It is unclear how false positives were filtered or how old links to deleted or modified videos were handled. Nor was it disclosed what share of the 6,454 videos is actually good.

There is also a systemic bias. HN has a narrow audience made up primarily of programmers, so the collection reflects their tastes and bubble. That is a benefit for developers, but it may be a drawback for other topics. Over time, videos that are already popular are also more likely to survive, while strong new work has not yet had time to accumulate mentions.

Conclusion

HackerTube is not a revolution, but a clear demonstration of what can be achieved by combining “a large archive of human recommendations plus an analyst model.” If you have a corpus of text in which people themselves identify what deserves attention, try turning it into a simple curated digest over a weekend. According to Karpathy's account, it now takes hours rather than weeks.