A recommendation system can’t show you a video it never admitted to the list.
That’s the choice I find interesting in HackerTube, Andrej Karpathy’s new video site built around recommendations in Hacker News comments. The familiar video grid is the visible part. The admission rule does a lot of the editorial work. 🍿
The live About page says a video needs links from at least three different people, with real commentary, across at least three different months to make the initial cut. It then uses AI to distinguish recommendations from other kinds of mentions.1, 2
Original explanatory diagram. The excluded video is a hypothetical example, not an audited HackerTube record.
Two videos walk into a filter
Imagine a brilliant new lecture. Someone posts it on HN with a thoughtful explanation. Nobody else has linked it yet.
Now imagine an older talk that people have brought up independently across several months.
Under the published initial rule, the new lecture hasn’t qualified. The older talk may have. That doesn’t tell me which lecture is better; it tells me what the system is built to notice. 💡
Repeated recommendations can be a useful signal for a durable reading or watching list. They also make brand-new material harder to admit. That trade-off starts before any personalized ranking.
If I were adapting this idea for fresh conference talks, I’d consider a separate lane for new arrivals. If I wanted an archive of recurring favorites, the repetition rule would be much closer to the task. Those are design choices, not improvements I’ve tested.
A link still needs interpretation 🔎
“This talk explains it well” and “this talk gets it wrong” can contain exactly the same URL.
Karpathy’s announcement describes using Claude to filter comments. The site’s About page spells out the distinction: recommending, citing, joking and criticizing are different stances. It also says the video cards and topic labels were written with AI.1, 2
That makes the comment-to-candidate step a classification problem. A polished summary doesn’t establish that the original commenter endorsed the video.
The useful detail in the live interface is that video pages link back to the HN comments. I’d follow those links when a card’s explanation matters. They let the reader inspect the evidence behind the summary instead of accepting the summary as evidence of itself. ✅
Ranking starts with a particular crowd
There’s another boundary I’d keep visible when borrowing the design: HN is a source community with its own interests. A list built from its comments can be valuable precisely because it has that perspective.
But “people here repeatedly recommended it” and “this is the best explanation for your audience” are different claims. A newcomer, a domain specialist and someone looking for a different language may need different pools of candidates before ranking even begins.
I’d check a few concrete things in any adaptation:
- Can readers see why a video qualified and return to its source comments?
- Are criticism and endorsement kept separate when the wording is ambiguous?
- Is there a way for useful new material to appear before it has a long mention history?
Those checks won’t turn curation into an objective quality score. They make the editorial choices easier to inspect. ⚠️
I checked the live product and its published explanation. I haven’t audited the complete data export, reproduced the classifier, or measured recommendation quality. The idea worth taking away is simpler: choose the candidate pool for the job you want the feed to do.
References and asset credit
- HackerTube, Andrej Karpathy: original announcement.
- About HackerTube: admission rules, comment classification and source links.
The figure is an original conceptual diagram. It contains no third-party thumbnails or copied screenshots.
AI disclosure
AI was used to draft the article.
Top comments (0)