# HackerTube HackerTube (https://hackertube.org) is a YouTube clone that only contains videos recommended in [[Hacker News]] comments. [[Andrej Karpathy]] built it on 31 October 2026, after a late-night YouTube session where the home page felt like "an enshittified wall of slop". His idea: "Let's write our own YouTube clone based on whatever videos have been recommended on Hacker News." The catalog is 6,454 videos recommended by HN users between 2007 and 2026, across 150 topics. The interface copies YouTube's: home feed, "Longs" (long videos), liked, history, watch later, categories, playlists. ## How it was built 1. Download every Hacker News comment from the `open-index/hacker-news` dataset on [[HuggingFace]] 2. Filter the comments, then run Claude Sonnet over them in a Claude Code workflow to drop spam. That leaves about 40,000 comments that recommend a YouTube video 3. Fetch the video metadata through the YouTube API 4. Wrap it all in a YouTube-like UI and deploy it to a VPS Karpathy's own hands-on part was creating the Google Cloud API key, booting the VPS and buying the domain. [[Claude Code]] did the rest, including deployment, security hardening and optimization. About two hours in total, done in between other Claude tasks. His comment: "I don't actually *do* anything resembling traditional engineering anymore." ## Design choices worth noting - **No accounts.** The "For You" recommendations work from state stored in your browser session. Nothing about you leaves your machine - **Middle-of-video thumbnails.** It shows the frame at 50% of the video, not the uploaded thumbnail. That one change removes the clickbait faces and red arrows - **Keyboard shortcuts** - **Data export.** You can export the dataset and give it to your own agent to build your own version ## Why it's interesting The recommendation signal is the opposite of YouTube's. YouTube ranks by predicted watch time ([[Algorithmic Curation]]); HackerTube starts from a person who took the time to write "you should watch this" in a technical discussion. That's a much smaller pool, but every item passed a human filter, and nobody's engagement metric shaped it. It's the same reason I still use Hacker News itself as my main tech source. It's also a clean example of [[Software for One]]: a tool built in an evening for one person's taste, using a public dataset, an LLM as the filter, and a coding agent as the engineer. The data export makes the point explicit. Karpathy isn't shipping a product; he's shipping a recipe. Swap the source (a subreddit, your [[Readwise]] highlights, a forum you trust) and you get your own feed. The limits are clear too. The catalog is frozen at the dataset snapshot, it only knows what HN's audience likes (lots of talks, lectures and engineering deep dives, very little else), and it depends on a Google API key to stay up. ## References - Site: https://hackertube.org - Andrej Karpathy, "HackerTube" (2026-10-31): https://karpathy.bearblog.dev/hacker-tube/ - Dataset: https://huggingface.co/datasets/open-index/hacker-news ## Related - [[Andrej Karpathy]] - [[Hacker News]] - [[Algorithmic Curation]] - [[Filter Bubbles]] - [[Attention Economy]] - [[Software for One]] - [[Vibe Coding]] - [[Claude Code]] - [[HuggingFace]] - [[2026-10-31 HackerTube - Karpathy rebuilt YouTube from Hacker News comments]]