# HackerTube's best feature might be the download **Summary:** HackerTube collects 6,454 videos discussed on Hacker News. I value its downloadable curation data because I can inspect the recommendations and change what makes the shortlist. - Canonical: https://markhuang.ai/news/hackertube-portable-curation - Language: en - Author: [Mark Huang](https://markhuang.ai/about) - Published: 2026-10-31 - Section: News - Tags: HackerTube, Andrej Karpathy, YouTube, Hacker News, Content Curation, Recommendation Systems - Source: [Andrej Karpathy](https://karpathy.bearblog.dev/hacker-tube/) - License: https://creativecommons.org/licenses/by-nc/4.0/ --- ![An open archive case holds glass cards showing a classroom, a scientific model, musical instruments, and a forest, with paper speech bubbles nearby](https://cdn.markhuang.ai/news/hackertube-portable-curation/hero.webp) *An editorial illustration of portable curation: taking the selected videos and the discussion behind them into a different tool.* Andrej Karpathy's HackerTube turns years of Hacker News video links into a YouTube-style browsing site. His [October 31 launch submission](https://news.ycombinator.com/item?id=50319510) lists 6,454 videos drawn from 39,834 comments since 2007. For someone looking for a talk worth watching, that offers a way in without first choosing a channel. I find the download more interesting than the homepage. HackerTube lets me borrow a community's taste, but its exported data also lets me question the selection and build a different way to browse it. That matters because the project's own documentation draws a distinction its pitch can blur: being discussed on Hacker News does not always mean being recommended. ## A repeated link still needs interpretation The [project's explanation](https://hackertube.org/about) describes a sensible filter. A video needs mentions with commentary from at least three different people across three different months to enter the collection. Claude then classifies the comments as recommendations, references, jokes, or criticism. The ranking rewards recommendations spread across different months, with extra weight for strong recommendations. I like that choice for finding enduring lectures. I am more inclined to make time for a talk people keep returning to than for a link that briefly becomes popular. But the first filter measures repeated discussion. The second depends on a model interpreting why someone shared the link. The [download's documentation](https://hackertube.org/download/DATA.md) makes the distinction explicit: about 25% of included videos were only cited and never recommended. That is not evidence that a quarter of the homepage consists of bad recommendations. It describes the underlying collection, which keeps several kinds of mentions and records recommendation counts separately. It does mean I would want an easy way to choose between "show me what HN discusses" and "show me what HN recommends." Both can be useful. If I am choosing a lecture for tonight, I would probably start with explicit recommendations. If I am researching a subject, citations and criticism might be exactly what I need. ## I am choosing a community's taste HackerTube's About page frames YouTube recommendations as tuned for watch time. I would be careful with that comparison. In its [2021 explanation of its recommendation system](https://blog.youtube/inside-youtube/on-youtubes-recommendation-system/), YouTube describes satisfaction surveys, sharing, likes, and dislikes alongside watch time. That account does not prove the feed serves every viewer well, but "watch time versus human judgment" is too simple a description of the difference. HackerTube also uses software to interpret human choices. Its advantage, to my eye, is that I can see whose choices form the starting point and inspect the comments behind a video. That makes the recommendation easier to question. It does not make the taste universal. The repeated-month rule also asks me to accept a tradeoff. A new video has not had time to accumulate recurring recommendations across months. I would use that signal to find material that has lasted, while keeping a separate route for recent discoveries. A [launch commenter](https://news.ycombinator.com/item?id=50319965) makes a useful distinction: recency can mean when a video was uploaded or when someone linked it again. I would want both dates. A renewed recommendation for an old talk could be more interesting than a newly uploaded video. ## The export lets me change the rules The [data download](https://hackertube.org/data) contains video cards, labeled comments, topic categories, and relationships between videos recommended by the same people. The documentation lists ODC-By 1.0 as the data license and requests attribution to HackerTube and the underlying Hacker News dataset. It also warns that the model's labels and summaries can be wrong. I can imagine using those files for a smaller tool. I would start with a weekly shortlist for a chosen topic, require an explicit recommendation, and keep a link to the comment that earned the video its place. I would rather get a short list I can act on than another feed to keep scrolling. The strongest objection is that nobody needs another viewing destination. A [commenter requesting YouTube playlists](https://news.ycombinator.com/item?id=50319987) argues that the curation is useful without a separate interface. I agree. Someone happy with their television's YouTube app should not need to change how they watch just to benefit from a better selection. The export gives a developer a way to answer that request, whether through another interface or a shortlist delivered elsewhere. There are limits: the download omits YouTube descriptions, view counts, durations, and upload dates, which the documentation explains how to fetch through the YouTube Data API. It is a curation dataset, not an offline video library. Karpathy also [says he has not published the server code](https://news.ycombinator.com/item?id=50320084). ## What would make me keep coming back I do not need HackerTube to prove that it has solved recommendations. I would want it to help me choose a worthwhile video and show enough of the reasoning that I can disagree. A model-written card is a convenient introduction; the linked discussion gives me something firmer to inspect. For me, the download is the feature that makes this experiment worth watching. I can accept Hacker News as one source of taste without accepting every inclusion or adopting another homepage. If the shortlist works better somewhere else, the project gives me a way to take it there. Source note: the submitted [Karpathy blog URL](https://karpathy.bearblog.dev/hacker-tube/) returned 404 when checked. The launch details above are corroborated by his Hacker News submission and HackerTube's project and data documentation.