How do I get started with the original HackerTube?
These steps apply to hackertube.org; other guide sections explain this Explorer site. Their search implementations, filters and saved lists differ.
- Open the original website and choose a subject using Search or Categories.
- Read the summary and HN context, then open the video.
- Add one filter to fit your viewing time.
- Save candidates with Watch later and bookmark important links separately.
How do I find the best videos for my needs?
Narrow the subject first, then try Best first. Hidden gems surfaces less-viewed material; Evergreen only excludes news-only items. Neither certifies quality. Recently recommended concerns discussion recency, not upload date. Longs helps find extended talks, but duration is not a measure of depth.
Read search and filters, saving videos and accounts and local state. Checked against the original feature explanation and client on November 3, 2026.
What HackerTube is
HackerTube is a browsing interface for videos that repeatedly appeared in Hacker News comments. Andrej Karpathy introduced the project on October 31, 2026 after asking Claude to build an alternative way to discover YouTube videos. The released snapshot contains 6,454 video records grouped into 150 categories across 16 broad areas.
The useful idea is simple: instead of starting from an engagement feed, start from links that people brought into real technical discussions over many years. The collection covers far more than programming. It includes mathematics, science, engineering, business, games, arts, history, health and practical skills.
The official project is at hackertube.org. HackerTube Explorer is an independent directory built from its public export and adds focused watch paths and a simpler explanation of the source signals.
The quickest way to start
If you already have a subject in mind, use the large search field on the homepage. Search checks video titles, channels, summaries, topic labels and tags. A specific phrase such as “linear algebra,” “compiler,” “backpropagation” or a channel name works better than a broad word such as “technology.”
If you want to browse, begin with the AI, machine learning and data guide, the software engineering guide or the mathematics and theory guide. Each one explains the territory before sending you into narrower category filters. You can then narrow the directory by format and level. The level field is a rough source label rather than an independently reviewed syllabus, so read the summary before deciding whether a video fits.
If you do not know what to search for, start with a watch path. The AI foundations, programming classics and mathematical intuition routes each order ten enduring videos around one learning goal. Topic guides help you choose a direction; watch paths give you a short sequence to finish.
How to read the Hacker News signals
A Hacker News appearance does not automatically mean a recommendation. The public export distinguishes positive recommendations from neutral references, criticism and jokes. About a quarter of the source records have no positive recommender in the retained labels. This site therefore opens with “Positive recommendations” selected.
The recommender count tells you how many distinct retained commenters positively pointed to the video. Recommendation months tell you how widely those endorsements were spread over time. HackerTube’s score adds distinct recommendation months and gives extra weight to months with strong recommendations. These are useful discovery signals, not objective quality scores.
Each watch page also explains why the video appears in the collection. Where the export includes a representative quote, the page links to its original Hacker News comment so you can read the surrounding discussion.
Searching and filtering videos
Search summaries and topics as well as titles
The directory index combines the title, channel, summary, Hacker News context, topic labels and tags. When you start searching, the browser loads the extended public text index and matches it locally; the phrase you enter is not sent to this site. Search uses keyword matching and a small reviewed Chinese–English glossary, rather than a semantic AI model. If the summary index cannot load, a notice identifies the narrower title, channel, topic and tag results.
Choose the right signal filter
Use Positive recommendations for the safest starting set. Evergreen only hides records labeled as dated or news. All mentions exposes the complete snapshot, including videos that were referenced neutrally or critically. The wider view is useful for research but should not be described as a list of universally endorsed videos.
Sort for the task
When you enter a search, the default order puts matches in titles and tags ahead of matches found only in summaries or broader topics, then uses HackerTube score to break ties. A small bilingual glossary supports terms such as linear algebra and backpropagation; it is not a translation of every video. Without a search, HackerTube score favors repeated recommendations over time. You can also sort by positive recommenders, recent discussion or title.
A worked search: backpropagation
Suppose you want to understand backpropagation rather than collect a long list of machine-learning videos. Search for “backpropagation,” keep Positive recommendations selected, and start with an Intro or General result whose summary explains gradients and error signals. The recommender count helps you compare repeated endorsements, but the title and summary should still match the question you are trying to answer.
From there, open the AI and machine-learning guide if you need the surrounding concepts, or use the AI foundations watch path when you want an ordered route from visual intuition into micrograd, Transformers and GPT. This is the practical difference between search, a topic guide and a watch path.
Watching and saving
Selecting a card opens a dedicated watch page with YouTube’s privacy-enhanced embedded player. Playback still depends on YouTube. A video may be unavailable because it was deleted, made private, blocked in your region or configured to reject embedding. In that case, try the direct YouTube link or return to the directory.
The Watch later button stores video IDs in your browser’s local storage. There is no account, synchronization or server-side profile. Clearing this site’s browser data also clears the watchlist. The same local list is available from the Watch later filter on the homepage.
What the data can and cannot tell you
The snapshot was exported on October 31, 2026. Its video summaries, categories, content types, levels and shelf-life labels were produced with language models. The sources and data page records the full attribution and known limitations, including labeling noise, duplicate uploads and the cap of twenty retained mentions per video.
YouTube descriptions, duration, publication dates, view counts, likes and current availability are not redistributed in the export. This directory does not invent those values. The source thumbnail is loaded from YouTube, while the title, channel and discussion context come from the published HackerTube snapshot.
Use the collection as a map rather than a verdict. The best final check is still to watch the video, consider its publication context and read the original discussion when the claim matters.
Choose your next step
Browse all 150 topics, choose an in-depth guide for AI and data, software or mathematics, or return to the complete video directory. For the original interface and its own recommendation features, visit HackerTube.