On 1 November 2026, Andrej Karpathy announced HackerTube (hackertube.org), built in about two hours with Claude. Tired of YouTube's recommendations, he built a site that shows only YouTube videos Hacker News commenters keep recommending, instead of what an engagement algorithm picks.
His X post passed 3.2 million views in 15 hours; the Hacker News post passed 2,800 points. The interesting part is what the build shows about working with AI agents — and its honest limits.
Note: the two hours are Karpathy's own estimate, not a benchmark, and nobody has reproduced the build. Blog post: 31 October; X post: 1 November.
What he actually built
Karpathy asked Claude Opus 5.5, in the Claude Code desktop app on a Mac (Medium effort as default), to filter 43 million Hacker News comments from 2007 to 2026 for YouTube video recommendations. Claude narrowed that to about 40,000 recommending comments — 39,834 rows — pointing to 6,454 distinct videos.
The rule was strict: a video qualifies only if at least three different people recommended it, with real commentary, in at least three different months. Claude judged whether each comment was recommending, citing, joking about or criticising the video. Teams of Claude Sonnet 5.5 agents, coordinated by Opus, wrote the comment labels, video cards and topic list — 150 topics — and videos are ranked by months recommended, not upvotes.
Karpathy's own work was three browser chores — a Google Cloud API key, a rented server (a VPS) and a domain name. Claude did the rest, including deployment: he told it to "update, secure, harden, deploy and optimize the code, and it did the whole thing."
The data is free as a 17.8 MB zip under the ODC-By 1.0 licence, with a DATA.md for agents: videos.jsonl (6,454 rows), mentions.jsonl (39,834 rows), related.jsonl (3,778 rows), and 150 topics. The site warns the labels were written by Claude, "so expect a few to be off."
Karpathy co-founded OpenAI, ran Tesla AI and Autopilot Vision (2017–2022), coined "vibe coding" (Feb 2025), and founded Eureka Labs.
Lesson 1. Start from a real annoyance
HackerTube did not start from a technology looking for a use. It started from Karpathy's frustration: YouTube's recommendations were not showing him what he wanted. A real annoyance is a clear test: does this make the annoying thing better?
Lesson 2. Let the crowd's recommendations pick, not an engagement algorithm
YouTube ranks by what keeps you watching. HackerTube ranks by how many months, over twenty years, different people recommended a video with real commentary — three people, in three months, before it appears at all. That filter does most of the work: 43 million comments became 6,454 videos. Curation beats generation — the agent did not create the taste, it found it buried in two decades of comments.
Lesson 3. Hand the agent the boring chores too
Karpathy's own contribution was three browser chores: an API key, a server and a domain. Everything else — updating, securing, hardening, deploying, optimising — went to Claude. Hand over the whole job, not just the fun part, or the human doing deployment stays the bottleneck.
Lesson 4. Use agent teams with clear roles
This was not one agent doing everything. Opus planned and coordinated, while teams of Sonnet 5.5 agents did the labelling — comment labels, video cards, topic list. One agent reading 43 million comments would lose track; splitting the work across specialised agents with a coordinator made the scale possible.
Lesson 5. Ship the data, not just the demo
HackerTube is a website, but the more durable thing Karpathy shipped is the dataset — 6,454 videos, 39,834 mentions, 3,778 related-video links and 150 topics, free to download. Shipping the data makes the work verifiable and reusable; the next person need not repeat the 43-million-comment filter. A demo impresses for a day; open data keeps working.
Key takeaways
- Karpathy built HackerTube in about two hours with Claude Opus 5.5 — his own estimate, not a benchmark.
- The site surfaces YouTube videos recommended by Hacker News commenters.
- Claude filtered 43 million comments to 39,834 recommending comments and 6,454 videos; a video needs three recommenders, in three different months.
- Karpathy's own work was an API key, a rented server and a domain; Claude handled deployment.
- The dataset is free, though some Claude-written labels will be off.
Conclusion
HackerTube is a small website with a big lesson: the most impressive AI builds are not the ones with the most code, but the ones with the clearest idea and the most honest accounting. Two hours is Karpathy's estimate, and the labels are imperfect — the site says so.
FAQ
What is HackerTube?
HackerTube is a website built by Andrej Karpathy in about two hours with Claude Opus 5.5. It shows only YouTube videos Hacker News commenters have repeatedly recommended, ranked by months recommended.
How did Claude choose the videos on HackerTube?
Claude filtered 43 million Hacker News comments from 2007 to 2026 to 39,834 recommending comments covering 6,454 distinct videos. A video qualifies only with three recommenders, with real commentary, in three different months.
Is the HackerTube data available to download?
Yes — free as a 17.8 MB zip under the ODC-By 1.0 licence, with a DATA.md for agents. The site warns some Claude-written labels may be off.
Sources
- https://cellcog.ai/blog/karpathy-hackertube/
- https://madrobot.blog/2026/11/03/karpathy-hackertube-youtube-built-with-claude/
- https://www.nextfor.ai/news/karpathy-hackertube-claude-build
Related reading on Daily Tech & AI:
Researched with AI assistance and reviewed for accuracy.
Follow on WhatsApp
Follow Daily Tech & AI on WhatsApp: https://www.whatsapp.com/channel/0029Vb9XzdHDJ6H4TFNAl61r
Comments
Post a Comment