Skip to Content
EasyTicket

Trying Karpathy's HackerTube, a YouTube Built With Words Alone

EasyTicket
Published:
Contents

Open the YouTube home page and you get an endless stream of videos that look a lot like the ones you watched yesterday. On October 31, 2026, Andrej Karpathy called this “an enshittified wall of slop” and posted a site on his blog. HackerTube is a website Karpathy released on October 31, 2026. It gathers 6,454 YouTube videos that people have recommended again and again in Hacker News comments over the years, and it shows them in a YouTube-style layout. It took about 2 hours to build, and the code was written not by a person but by Claude Code.

The raw material for HackerTube is comments on Hacker News (an English-language news community where developers gather). From all the comments accumulated from 2007 to 2026, the project pulled out YouTube links, then narrowed them to about 40,000 comments (exactly 39,834) that recommend a video. That produced 6,454 videos, which were sorted into 150 topics and 16 categories.

The selection rule shapes the character of the site. According to the About page, a video stays only if three different people linked it with an explanation, across at least three different months. Rankings also depend not on view counts but on “how many months it was recommended over.” A video that someone keeps bringing back year after year rises above one that went viral once.

HackerTube About page. Four numbers (6,454 videos, 39,834 comments, 150 topics, 2007 to 2026) and an explanation of how it was made
The About page starts with four numbers and the criteria used to pick videos.

Why It Spread So Widely

Start with the numbers. The Hacker News post Karpathy submitted himself had 2,882 points and 640 comments as of November 6, 2026. On Hacker News, that is a post that defines its week. In Korea, GeekNews (a Korean tech news site) covered it on November 1, and Japan’s GIGAZINE and India’s Business Today also ran stories.

Bar chart comparing Hacker News post points. HackerTube 2,882 points, HackerBooks 1,860 points
Hacker News points as of November 6, 2026. The original is about 1,000 points ahead of the book version that followed.

The sources point to two reasons. In Hacker News comments, the most common praise was that the results are far more relevant than the YouTube home page. Business Today highlighted that you can use it without an account, and that its thumbnails use a frame from the middle of the video. Instead of the exaggerated faces typical of YouTube thumbnails, you see the frame at the 50% mark, so the clickbait imagery disappears.

EasyTicket sees one more reason: “who made it, and how” was as big a topic as the site itself. A well-known AI researcher said he handed over only a server address and had AI do everything else, so people who came to look at the site and people who came to see how it was made arrived at the same time. There was also a moment on launch day when someone asked in a comment for sorting by newest, and Karpathy replied in the same thread that he had added sorting and filters. It showed how fast development by talking can move.

Two Hours of Asking Claude Code, in Words Only

The steps written on Karpathy’s blog are short.

  1. Download the entire Hacker News comment dataset published on Hugging Face
  2. Use Claude Sonnet to filter out spam and unrelated comments
  3. Keep only the comments that recommend YouTube videos
  4. Fetch video information (title, view count and so on) from YouTube
  5. Build a YouTube-like interface
  6. Put it online with a domain and a server

The tool was the Claude Code desktop app for Mac, and the default model was Opus 5.5 (Medium effort). The human did only three things by hand: creating a Google Cloud API key to fetch YouTube data, spinning up one server (a VPS), and buying a domain. After he handed over the server IP, Claude reportedly handled deployment, security and speed tuning. In Karpathy’s words, “I only talk to Claude and have it do everything.”

The hint this offers to anyone who wants to try vibe coding is clear. The hard part was not the code but the standard for “what to keep.” Three people, three months, a link with an explanation. That one-line rule created the quality of the 6,454 videos.

How to Use HackerTube

You don’t need to sign up. Open the address and a grid of videos appears right away. On a computer the screen filled in after about 3 seconds, and everything from the left menu (Home, Longs, Liked, History, Watch later, Categories, Playlists, Settings) to the search bar at the top copies YouTube almost exactly. Likes and watch history are saved in your browser instead of an account.

HackerTube home screen. Left menu, search bar at the top, and a three-column grid of thumbnails
The home screen is laid out almost exactly like YouTube. The order of videos changes every time you enter.

Once you go in, you see that the first row already mixes eras. A TED talk from around 2012, a Linux conference presentation and a RustConf 2018 keynote sit side by side, and a video uploaded 9 months ago appears next to one that is 20 years old. Nearly half of the talks run longer than an hour. The home order changes each time you reopen it, so you never see the same home page twice.

The range of topics was a surprise. I expected only programming videos because it comes from Hacker News, but when you open Categories, the largest category is art, comedy, music and film (1,731 videos). Software engineering (1,604) is second, and politics, history and society (1,425) follows right behind. Stand-up comedy and clips from John Oliver’s show also turned up on the home page.

HackerTube categories screen. Cards for 16 categories and the number of videos per subtopic
The largest category is not programming but art, comedy, music and film (1,731 videos).

My favorite button is Hidden gems, at the top of Categories. Press it and conference talks from RubyConf, PyCon Russia, elm-conf and others line up with view counts in the 20,000 to 80,000 range. These are videos the YouTube algorithm would almost never put on your first screen. You can also send the whole list to YouTube with “Play all on YouTube” at the upper right.

HackerTube Hidden gems list. Developer conference talks with 20,000 to 80,000 views
Hidden gems collects conference talks that have few views but have been recommended for a long time.

The real difference of this site shows when you open a single video. Open the elm-conf talk “Making Impossible States Impossible” and under the title you see “9 HN recommenders · recommended in 13 months, 2017–2026.” It means nine people recommended this video over 13 months spread across nine years. Below that are a summary, a one-line note on why Hacker News people treasure the video, and topic tags. Scroll further and 19 Hacker News comments follow in their original text, each with a label such as “Strongly recommends.”

HackerTube video detail screen. Number of recommenders, AI summary, topic tags, and a list of related videos on the right
Under the video title, the site shows 9 recommenders and a record of recommendations across 13 months.

When you can’t decide whether to watch, being able to read real recommendations from strangers is the biggest difference from YouTube. One caution: as the About page explains, the summaries and labels such as “evergreen” were written by Claude, so it helps to treat the comments below as the place where the human voices are. The video window is a YouTube embed, so playback still runs through the YouTube player.

HackerBooks and HackerTV Followed

Within days of HackerTube taking off, variations using the same approach were posted on Hacker News. The one that drew the biggest response was the book version, HackerBooks. Its creator said Claude (Opus 5.5) wrote almost all of the data processing and site code, and as of November 6 it had 1,860 points and 307 comments. HackerTV, which turns the same data into a 24-hour TV channel, also appeared, and GeekNews covered it on November 4.

HackerTubeHackerBooksHackerTV
What it collects6,454 YouTube videosMore than 15,000 books mentioned by two or more peopleYouTube videos recommended multiple times on HN
How it displaysYouTube-style grid and detail screensGrid of book covers, publication-era filter24-hour schedule by channel
Recommendation judgmentClaude classifies comments as recommendation, citation, joke or criticismClaude classifies each mention as recommend, neutral or criticismNot confirmed
Hacker News response (Nov. 6)2,882 points, 640 comments1,860 points, 307 commentsNot confirmed

Go into HackerBooks and “Today’s shelf” holds 15,704 books, with each book showing its recommendation count and total mention count, as in “218 recommend · 294.” Its filters are denser than the video site’s, and the large covers make browsing easy. The topic ranking is fiction and literature (3,366 books), then politics and society, then biography, with programming in fourth place, which resembles HackerTube. It suggests Hacker News people recommend quite a lot outside of programming.

HackerBooks home screen. Grid of book covers with recommendation and mention counts for each book
HackerBooks, the book version, shows recommendation and mention counts together on a cover grid.

HackerBooks first counted books as “recommended,” then changed the wording to “mentioned by” after people pointed out that critical mentions were mixed in. It shows that separating recommendation from mention is the heart of this kind of site. If you want to pick a video, HackerTube is the right one. If you want to find your next book, go with HackerBooks.

The Downsides: English and an Old List

The first is language. The summaries and reasons for recommendations are all in English, and I didn’t see a setting to switch to Korean. Most videos are English talks that are hard to follow without subtitles, so if listening to English isn’t comfortable, the practical route is to send the video you picked to YouTube and turn on auto-captions.

The second is the age of the list. Because videos recommended repeatedly over many months rise to the top, you often see videos more than 10 years old. Business Today also reported that some users felt the list was dated. It is not the right site for someone looking for recent AI videos.

Third, the home page also mixes in videos on political and social issues. If you want to see only programming videos, you have to go through Categories or Filters first. Hacker News commenters also worried that as the site gets more popular, spam and manipulation aimed at recommendation comments will grow, and about who looks after the people who make the videos.

Frequently Asked Questions

Do I have to sign up to use it?

No. You can watch right away without an account. Likes, watch later and watch history are saved in the browser you use, so on another device they start over from scratch.

Can I watch on my phone?

Yes. Open it in a phone browser and the menu folds behind a button, with one large thumbnail per row. There is no separate app.

Can I build a similar site myself?

On the About page, under “Fork it,” you can download a 17.8MB compressed file containing the video list, labeled comments and topic classification. The page says it also includes instructions to give this file to an AI agent and tell it what you want, plus a guide written for agents. Given that variations like HackerBooks appeared within days, the barrier is fairly low. However, the site’s source code has not been released.

Will new videos keep being added?

A question about whether there are plans to re-collect data periodically came up in the GeekNews comments, but as of November 6, 2026, no update schedule has been confirmed. The current list was made by going through Hacker News comments from 2007 to 2026 once.

Sources 6
  1. HackerTube (karpathy.bearblog.dev)
  2. HackerTube: YouTube built from videos recommended by HN (Hacker News)
  3. HackerTube - Karpathy가 만든 Hacker News 추천 영상 모음 (GeekNews)
  4. 'HackerTube' displays 6,454 videos collected from 39,834 comments on Hacker News (GIGAZINE)
  5. Tired of YouTube recommendations? Andrej Karpathy builds HackerTube with 6,454 videos to explore (Business Today)
  6. Hacker Books' Top 10 Ranking (GIGAZINE)

AI helped with research and the first draft; the operator tried the subject firsthand, checked and edited this article before publishing. Let us know if you spot a mistake.