HackerTube is a video discovery site built from YouTube links that people have shared on Hacker News over the years.
Andrej Karpathy created it after becoming frustrated with his YouTube homepage. He wanted an evening’s viewing to begin with something worth watching, rather than another round of exaggerated thumbnails and familiar suggestions.
He says the project took about two hours with Claude Code. The result raises two useful questions: how much of a small product can a coding agent build, and what happens when a community’s recommendations become the foundation of a video browser? Karpathy’s announcement
To understand the experiment, it helps to separate three things: the people who recommended the videos, the AI that organized those recommendations, and the interface through which viewers discover them.
1. Why build another way to browse YouTube?
Imagine asking a group of people you trust for one good lecture. Someone remembers a talk that helped them understand a difficult subject. Another person recommends the same talk months later, in a different discussion.
That repetition tells you something. The video has remained useful across more than one conversation.
Now imagine those recommendations scattered across almost twenty years of discussion. The knowledge exists, but finding it requires reading an archive.
HackerTube turns that archive into a library you can browse.
In his account of the project, Karpathy describes opening YouTube late on a Friday and disliking what he saw. The next morning, he asked Claude to build a YouTube-style site around videos recommended on Hacker News. The original site is hackertube.org. Karpathy’s project notes
The useful product idea is small and specific: keep access to the existing videos, and create a discovery experience around a source of judgment you prefer.
2. What “built in two hours” actually means
Karpathy says he used the Claude Code desktop app for Mac, with Opus 5.5 at Medium effort. He directed the work through conversation, while Claude handled the implementation.
He also identifies three things he did manually:
- Create a Google Cloud API key for retrieving YouTube video metadata.
- Start a virtual private server, or VPS, to serve the site.
- Buy the domain name.
After the server was available, he gave Claude its IP address and asked it to deploy, secure, harden, and optimize the application. These are his reported actions and timing; the account does not provide an independent security audit or a reproducible two-hour benchmark. Karpathy’s account of the build
This explains why the project is more interesting than a quickly generated homepage. It includes a data collection process, interpretation of comments, search, recommendations, and a deployed interface.
The scope also helps explain the speed. HackerTube uses an existing discussion archive and existing YouTube videos. Its main job is to connect those resources in a useful way.
3. How the collection was built
The headline numbers describe different stages of the process. Karpathy’s announcement refers to roughly 43 million HN comments in the starting archive. The site’s About page describes 39,834 labeled comments and 6,454 videos, organized into 150 topics across 16 areas. The larger number is the search space; the smaller numbers describe the resulting collection. Announcement, HackerTube’s methodology
The official explanation breaks the work into a practical sequence.
First, find YouTube links. Scan the public HN comment archive from 2007–2026. This produces well over a hundred thousand distinct candidate videos.
Second, look for repeated interest. The site says a candidate needs links from at least three different people, with substantive commentary, across at least three different months. This reduces the candidates to roughly 7,000 videos.
Third, interpret the comments. A link can recommend a video, criticize it, cite it as evidence, or use it as a joke. Claude reads the comment with its HN story title and labels its purpose.
Fourth, create usable video records. Combine the discussion with YouTube details to produce summaries, topics, and explanations of why people found a video worth sharing.
Finally, rank for lasting interest. The site emphasizes the number of separate months in which a video received recommendations. Repeated recommendations over time count more than one burst of attention. Official description of the pipeline
Karpathy later clarified that Claude first reduced the archive using URLs, regular expressions, and spam removal. Only the remaining material went to Sonnet through a Claude Code Workflow. The headline therefore does not mean a language model individually read all 43 million comments. Karpathy’s clarification
The order matters. A program can cheaply recognize a URL. Understanding whether a sentence endorses that URL requires more interpretation.
This is a useful pattern for similar projects: narrow the material with straightforward rules, then use a language model on the smaller set that needs judgment. The model’s output still needs scrutiny, especially where sarcasm, criticism, or ambiguous recommendations are involved.
4. Human recommendations, organized by AI
Calling HackerTube “human curation” needs a little precision.
HN users supplied the original links and commentary. AI classified those comments and wrote the browsing material. Software then ranks and personalizes the collection.
The official homepage mixes established favorites, less widely watched discoveries, recent favorites, and topics the viewer has not explored. Likes can influence subsequent suggestions. Video pages also link back to the original HN comments, so viewers can inspect the context themselves. HackerTube’s feature explanation
That combination is the experiment: human recommendations provide the underlying signal, while AI makes a large, scattered archive easier to use.
Does this prove that people recommend better videos than algorithms?
Karpathy reports a much better viewing experience for himself. That is a meaningful reason to build the product, but it does not establish a result for every viewer. Hacker News has its own interests, blind spots, and community biases. A library that fits a programmer’s evening may fit someone else poorly. Karpathy’s assessment
The narrower conclusion is still valuable: a community whose interests resemble yours can provide a strong starting point for discovery.
5. Small choices that change the experience
Karpathy also changed the thumbnails. His announcement says HackerTube uses a frame from the middle of a video instead of the uploader’s selected promotional image. That choice gives viewers a different first impression of the same underlying material. Thumbnail design
The original site requires no account. Likes, history, and saved videos stay in the browser. It also includes a Longs section, topics, playlists, and keyboard shortcuts. Local preferences can still support useful personalization, although they do not automatically become a synchronized account across devices. Features and privacy
None of these choices is complicated to describe. Together, they express a coherent preference: help someone choose a worthwhile video with less ceremony.
6. An independent alternative: hackertubes.com
There is also hackertubes.com, an independent browser built from the original project’s exported data. Its attribution explicitly says it is unaffiliated with the original HackerTube project.
Its design has a Web 1.0 sensibility: an understated wordmark, plain controls, a directory of videos, and explicit pagination. The page feels closer to a collection someone assembled for curious readers than a constantly moving social feed.
The directory supports five languages: English, Simplified Chinese, Traditional Chinese, Korean, and Japanese. These localized interfaces make the collection accessible to more readers, while the videos retain their original languages.
The site offers topic and duration filters, a daily mix, an HN favorites order, and browser-local watched and watch-later lists.
Its daily mix reshuffles the existing archive; it does not promise newly uploaded videos every day. The directory displayed 4,616 videos when checked on November 2, 2026. Its data notes say it excludes news and entries without a recommendation, so its collection differs from the original 6,454-video export. Independent directory and data notes
For readers who prefer a quieter directory and a choice of interface language, it is another way into the collection.
7. Why the export matters
HackerTube makes its underlying collection downloadable. The official About page lists video records, labeled mentions, topic definitions, related-video data, search embeddings, and a DATA.md guide intended to help an agent work with the files.
The data is offered under ODC-By 1.0 with attribution. YouTube’s own metadata is excluded from the export; builders can retrieve it separately using the video IDs. Official export documentation
This gives the experiment a life beyond its first interface.
A useful collection could support a weekly reading-and-viewing list, a subject-specific study queue, or a simpler browser in another language. These are possible applications of the export, rather than claims that Karpathy built all of them.
The independent directory illustrates one such possibility: the same source material can support a different pace and presentation.
8. What the experiment means
HackerTube brings together two resources: years of people pointing one another toward useful material, and a coding agent that can turn those scattered pointers into a working product quickly.
The two-hour claim concerns the software-building effort. The community’s accumulated judgment took years to form. AI makes that judgment easier to retrieve and organize; it also introduces classification and summary errors that viewers should be able to check against the original discussion.
For builders, the practical lesson is to look for a narrow frustration and a good existing source of information. A small application can make that source dramatically more useful without needing to create the content itself.
For viewers, the invitation is simpler: choose whose recommendations deserve your attention, then choose an interface that helps you use them.
Top comments (0)