A video can be excellent and still never appear on your screen. That is the problem HackerTube tries to solve, but not by making a better version of YouTube's recommendation model. It changes the source of the recommendations.
HackerTube starts with videos that people recommended in Hacker News comments over almost two decades. YouTube's homepage, by contrast, is personalized around your own viewing history and other signals. One system asks, in effect, "What has this community found worth sharing?" The other asks, "What might this particular viewer want next?"
That difference makes HackerTube an unusually clear comparison of two discovery strategies: a shared human-curated corpus versus an individualized platform feed. Neither is automatically better. Each has a different blind spot, and the choice depends on whether you want a common catalog of recommendations or a feed tuned to your own habits.
I have not audited HackerTube's dataset or run a controlled comparison against YouTube. This is an analysis of the project creator's account, the site's stated features, YouTube's own description of its recommendation system, and the developer discussion around the launch.
What HackerTube changes
Andrej Karpathy introduced HackerTube on October 31, 2026. He says it draws from 39,834 Hacker News comments, narrowed to roughly 40,000 comments that recommend videos, and contains 6,454 videos across 150 topics from 2007 through 2026. Those are creator-reported figures, not an independent audit. The project presents those videos through a familiar video-browsing interface. Karpathy's launch post describes the collection and workflow; the Hacker News launch thread summarizes the catalog.
The source of its signal is the important part. A Hacker News user had to encounter a video, decide it was worth mentioning, and place it in a discussion. That is not a formal quality review, and a comment recommendation can be a joke, a disagreement, or a niche reference. But it is a trace of human selection that exists outside your own watch history.
The site's design also makes a few deliberate choices. Karpathy says it uses a frame from the middle of a video instead of the creator-uploaded thumbnail, offers recommendations without accounts, and keeps browsing state in the local browser session. He also says users can export the site's data and use it to build a different interface. These are descriptions from the creator, not guarantees about long-term availability or a substitute for inspecting the implementation. The launch post lists these features.
The project is not a new video host. The catalog points people toward YouTube videos. The discovery layer changes; the underlying video source does not. That distinction matters when assessing what the project can and cannot control.
What YouTube optimizes for
YouTube describes its recommendations as a personalized system. Its official explanation says the homepage primarily uses watch history and considers signals including search history, subscriptions, likes, dislikes, feedback such as "Not interested," and satisfaction surveys. The company says the system compares your viewing habits with those of people with similar habits.
That design can be powerful. If you have specific interests, a long viewing history, and channels you return to, the system can make useful suggestions that a general-purpose catalog would miss. It can also react to your feedback and viewing patterns rather than relying on what an outside community liked years ago.
The system's complexity is not proof that it serves every individual well. YouTube's explanation describes the signals it uses and its stated goal; it does not establish that every viewer experiences the feed as satisfying. The HN launch discussion contains both reactions: one commenter said the unpersonalized feed felt unusually close to their interests, while another said HackerTube resembled an HN-shaped algorithm and included videos they did not want. Those are personal reports, not representative user research. The thread is useful precisely because it shows that a community-derived feed can feel refreshing to one person and narrow to another.
HackerTube vs. YouTube: the practical trade-offs
Discovery signal: HackerTube uses historical links people chose to recommend in Hacker News discussions. YouTube uses personal and contextual signals, including viewing and search history. If you want to step outside your established viewing pattern, HackerTube's external corpus is attractive. If you want suggestions based on what you actually watch, YouTube has the more direct feedback loop. HackerTube's creator describes its source; YouTube describes its signals.
Breadth versus fit: HackerTube's reported 6,454 videos and 150 topics are a finite, shared collection. Its selection may surface a talk or documentary you would not find through your usual recommendations, but the collection reflects what was linked in one community. YouTube's catalog and personalization can cover far more viewing contexts, but a personalized feed may keep reinforcing patterns already present in your history. The latter is a possible trade-off, not proof that the system inevitably creates a filter bubble.
Freshness: HackerTube's corpus reaches back to 2007. A historical recommendation can remain useful, especially for foundational talks, but a decades-long window can also make old material prominent. YouTube can use current behavior and context to suggest recent uploads, though the results are still ranked for each viewer. HackerTube has since added a "Recently recommended" sort, according to a creator update in the launch thread. That helps distinguish a new HN mention from an old upload, but it does not make every video itself new.
Thumbnail incentives: HackerTube's middle-frame thumbnail is meant to avoid relying on the creator's selected thumbnail. It changes the preview, not the video's content, title, or reputation. YouTube's recommendation system considers whether viewers choose to watch, stay, and report satisfaction, according to its official overview. The two interfaces therefore expose different cues, but HackerTube has not demonstrated that one thumbnail strategy produces better viewing outcomes.
Personalization and privacy: HackerTube says it does not require accounts and stores browsing state in the local browser session. YouTube says its recommendations use account-associated signals such as watch history and subscriptions, and it provides controls to pause or delete that history. The right choice depends on the experience you want: minimal personalization or a feed that adapts to your behavior. Check each service's current settings and privacy explanation before treating these descriptions as a complete privacy comparison. HackerTube's creator post and YouTube's explanation describe the respective approaches.
Control and escape hatches: YouTube offers controls including "Not interested," "Don't recommend channel," and watch-history management. HackerTube's exportable catalog offers a different kind of control: a dataset someone can inspect or reuse to create another interface. One adjusts a personal feed; the other allows people to fork the collection's presentation, subject to the data and service remaining available. YouTube's help page lists its feedback signals; HackerTube's About link is linked from the launch discussion.
A recommendation feed is a product decision
The deeper lesson is not that one algorithm is honest and the other is manipulative. The two systems answer different questions because they start from different objectives and evidence.
HackerTube uses a community's past recommendations as a shared taste map. That map can help people discover material outside their own viewing history. But it inherits the limits of that community: who participates, which links get posted, which discussions survive, and what a commenter meant by sharing a video. A popular recommendation can be a useful clue, not a quality certificate.
YouTube's system personalizes suggestions from a broader stream of signals. That can make discovery more relevant to an individual, but it also means the experience is shaped by a platform's ranking choices and the data available about that viewer. YouTube says it uses satisfaction surveys alongside watch behavior. That is a meaningful distinction from a simplistic claim that the system only maximizes time watched. The actual outcome for any viewer still needs to be judged from their experience, not inferred from the system description alone.
HackerTube's smaller scale is a feature for some people and a limitation for others. A curated catalog makes the selection legible: the underlying videos were surfaced in public discussions, and the project reports the collection size and topics. But the process that filtered comments and grouped content is not the same as a transparent editorial review. Karpathy says AI models helped filter and classify the comments; that does not make the resulting list neutral or objectively high quality. His post explains the broad workflow, while the HN discussion surfaces questions about curation and coverage.
Which one should you use?
Choose HackerTube when you want a finite, community-shaped browsing session, especially if you are curious about technical talks and videos recommended in Hacker News discussions. Treat the catalog as a discovery index, then decide for yourself whether each video deserves your time.
Choose YouTube when you want a personalized feed, recent uploads from channels you follow, or recommendations that can adapt to your interests and feedback. If the homepage has drifted away from what you want, review your watch history, subscription list, and recommendation controls before deciding the whole system is useless. YouTube's own recommendation controls guide explains how to manage those inputs.
You can also combine them. Use HackerTube as a source of candidates, and YouTube as the player and subscription tool. That avoids pretending either source is universally superior. It also gives you a practical test: do you find videos worth watching that you would not have found through your normal feed?
A five-minute evaluation checklist
- Pick a goal. Are you looking for a specific topic, a new creator, a long-form talk, or something recent? A discovery tool can only be judged against a concrete job.
- Set a short session. Browse for ten minutes and save candidates instead of letting autoplay decide the length of the experiment.
- Track useful finds. Count videos you actually want to watch, not just thumbnails you click. Note whether you already knew the creator or topic.
- Check provenance. For a HackerTube result, inspect the discussion context when available. A recommendation's reason can be more informative than the fact that it appeared in a list.
- Compare the cost. Consider time spent, interruptions, account requirements, and how easy it is to leave with a short watchlist.
- Repeat once. A single session is anecdotal. Try a different topic or day before deciding one approach consistently works better for you.
The best feed is not necessarily the one with the most advanced model. It is the one whose recommendations match your intention without turning every visit into an unplanned session.
HackerTube is an interesting counterexample to the assumption that discovery must begin with a platform's own personalization. It asks whether a public archive of human recommendations can be useful enough to browse on its own. YouTube answers a different question: given the signals available about this viewer, what might they want next? For some viewers, switching between those answers may be more useful than choosing a permanent winner.
I write about developer tools, AI systems, and practical software engineering. Subscribe if you want more comparisons that focus on how tools behave, not just how they are announced.
Have you tried HackerTube or another community-curated video catalog? What did it help you find that your usual feed missed?
Top comments (0)