# Karpathy's HackerTube: How Claude Built It in 2 Hours

> Andrej Karpathy had Claude filter 43M Hacker News comments down to 6,454 recommended videos in about two hours. How HackerTube works.

- Author: Nitish Garg, Founder & CEO, CellCog
- Published: 2026-11-02
- Canonical (HTML): https://cellcog.ai/blog/karpathy-hackertube/
- Section: Guides / Choosing a platform
- Publisher: CellCog (https://cellcog.ai), the AI employee platform. Blog index for agents: https://cellcog.ai/blog/llms.txt

## Key points

- On November 1, 2026 Andrej Karpathy launched HackerTube, a site of the YouTube videos Hacker News keeps recommending, built in about two hours with Claude Opus 5.5.
- Claude filtered 43M comments over 20 years of Hacker News down to 6,454 videos.
- A video counts only if at least three people linked it with real commentary in at least three different months, and it ranks by months recommended.
- Opus built the site and pipeline; teams of Claude Sonnet 5.5 agents it coordinated wrote the comment labels and video cards.
- The data is a free 17.8 MB zip under ODC-By 1.0 with a DATA.md written for agents, and others are already forking it.

## At a glance

- **What is HackerTube?** A free site of 6,454 YouTube videos that Hacker News commenters kept recommending from 2007 to 2026, launched by Andrej Karpathy on November 1, 2026.
- **How was it built?** Claude Code with Opus 5.5, in about two hours by Karpathy's account; he only made an API key, rented a server and bought the domain.
- **Can I build my own?** Yes. The data is a 17.8 MB zip under ODC-By 1.0, with a DATA.md written for AI agents.

On November 1, 2026, Andrej Karpathy posted a weekend project: [HackerTube](https://hackertube.org), a video site that only recommends the YouTube talks Hacker News commenters keep pointing each other to. He wrote that he "had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended" in about two hours with Opus 5.5. [His post](https://x.com/karpathy/status/2116957135661945283) passed 3.2 million views within 15 hours.

His own part of the build was three browser chores: a Google Cloud API key for YouTube metadata, a rented server and a domain name. Claude Code did the rest, including the deploy: "Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing."

## How HackerTube picks a video

![Animated diagram: a stack of Hacker News comment bubbles passes a gate of 3+ people and 3+ months to about 7,000 videos, a Claude card sorts each comment into Recommend, Cite, Joke or Criticize, and only Recommend flows on to a shelf of 6,454 videos ranked by months](https://cellcog.ai/blog/media/karpathy-hackertube/funnel.svg)
*A link is not a recommendation: two filters turn 43M comments into 6,454 ranked videos*

The site's [about page](https://hackertube.org/about) lays out the pipeline. The useful idea is the filter: a link on Hacker News is not the same as a recommendation.

*Table: How HackerTube was built, per its about page (read November 2, 2026)*

| Step | What happened | Result |
|---|---|---|
| Scan | Every HN comment from 2007 to 2026 searched for YouTube links | Well over 100,000 different videos |
| Keep | Linked with real commentary by at least three people, in at least three different months | Roughly 7,000 videos |
| Read | Claude read each comment with its story title and judged the stance: recommend, cite, joke or criticize | 39,834 labeled comments |
| Describe | Claude wrote a card per video: summary, why HN likes it, type, whether it aged well | 6,454 video cards |
| Rank | Standing comes from how many separate months people recommended it | Lasting picks beat one viral spike |

The ranking rule is the part worth copying: "Being linked again and again over the years counts for more than one viral spike." It is the opposite of a watch-time feed, which the site says is "very good at finding the next thing you'll click, and not so good at finding the talk you'll still be thinking about next year."

## Who did the work

Karpathy's setup was the Claude Code desktop app for Mac, with Opus 5.5 on Medium effort as his default. The site's own credits split the job in two: "HackerTube was built by Claude Opus 5.5, Anthropic's AI model, working in Claude Code: the data pipeline, the search and recommendations, and this site."

The reading was delegated: "The comment labels, video cards and topic list were written by teams of Claude Sonnet 5.5 agents that Claude Opus coordinated." Reading tens of thousands of comments is the expensive, repetitive part. One coordinating model handing it to teams of smaller ones is how that kind of job stays inside a weekend.

## What is in the download

Everything behind the site is a 17.8 MB zip on the [data page](https://hackertube.org/data), exported October 31, 2026, under ODC-By 1.0: "use it for anything, just credit HackerTube."

*Table: What is in the HackerTube download (hackertube.org/data, read November 2, 2026)*

| File | What it holds | Rows | Size |
|---|---|---|---|
| videos.jsonl | One line per video: title, channel, summary, why HN likes it, topics, score | 6,454 | 8.6 MB |
| mentions.jsonl | One line per labeled HN comment: stance, quote, text | 39,834 | 34.7 MB |
| related.jsonl | Videos recommended by the same HN people, counts only | 3,778 | 1.5 MB |
| taxonomy.json | Topic areas and categories | 150 | 44 KB |

YouTube's own metadata (descriptions, view counts, durations) is not in the zip, because YouTube's API terms do not allow redistributing it. Every record carries the video id, and the data page says a free YouTube API key fetches the rest in about 130 requests.

## Built to be forked by an agent

The zip is meant to be handed to an agent, not opened by hand: "The zip includes a DATA.md written for agents: how the files fit together, the gotchas, example records, and how to walk you through getting YouTube metadata if you want it." Karpathy's own post ends the same way: point your Claude at the data and ask for your own version.

People did. Within seven hours, [elvis (@omarsar0)](https://x.com/omarsar0/status/2117053859516285354) said he "took his prompt and similar data sources and asked my agent to build the arXiv papers version of it", and published [Hacker Papers](https://academy.dair.ai/papers/collections/hacker-papers) with more than 1,000 papers.

## What the post does not show

- **One weekend project, self-reported.** The two hours are Karpathy's estimate; there are no token or cost figures.
- **Labels can be wrong.** The data page says it plainly: "The labels, summaries and cards were written by Claude, so expect a few to be off."
- **Not a YouTube replacement.** Videos play through YouTube's privacy-enhanced embed, and YouTube's own metadata stays with YouTube.
- **No accounts.** Likes, history and Watch later live only in your browser, so nothing follows you between devices.

## Our read

This is the second post in a day where one person states the outcome and Claude Code picks the tools: Pieter Levels [had it edit his vlog](https://cellcog.ai/blog/claude-code-vlog-editing/) the same Sunday. The human part shrinks to what needs a login, a card or a judgment. The detail we would point at is the split inside HackerTube: Opus coordinating teams of a smaller model to read 39,834 comments. That is how agent teams take on reading work at scale.

CellCog's AI employees run on Claude Opus 5.5, in their own workspace, with memory that carries from one task to the next.

## Sources

- X: [@karpathy, November 1, 2026](https://x.com/karpathy/status/2116957135661945283) and [@omarsar0](https://x.com/omarsar0/status/2117053859516285354). Read directly; times from the post ids.
- HackerTube: [about](https://hackertube.org/about) and [data](https://hackertube.org/data) pages, read November 2, 2026.
- The underlying archive: [open-index/hacker-news on Hugging Face](https://huggingface.co/datasets/open-index/hacker-news).

## FAQ

**How does HackerTube choose videos?**

A video makes the cut if at least three different people linked it with real commentary in at least three different months, then ranks by how many separate months it was recommended.

**Did Claude really read the comments?**

Yes. The site says Claude read each comment with its story title and labeled whether it recommended, cited, joked about or criticized the video.

**Which Claude models built it?**

The site credits Claude Opus 5.5 in Claude Code for the pipeline, search and site, and teams of Claude Sonnet 5.5 agents coordinated by Opus for the labels and cards.

**Is the data free to use?**

Yes, under ODC-By 1.0: use it for anything and credit HackerTube and the open-index/hacker-news dataset. YouTube's own metadata is not included.

**Does HackerTube need an account?**

No. Likes, history and Watch later are stored only in your browser.

## Related

- [Pieter Levels Let Claude Code Edit His Vlog: The Pipeline](https://cellcog.ai/blog/claude-code-vlog-editing/index.md)
- [Karpathy's AI Funnel: Only 5,000 See the Frontier](https://cellcog.ai/blog/karpathy-llm-understanding-funnel/index.md)

## The AI employee for this read

[AI Data Analyst](https://cellcog.ai/ai-employees/ai-data-analyst): I built this page. For what it covers, hire a data analyst: it pulls the numbers, checks them twice and sends the chart with its source.

---

Markdown alternate of https://cellcog.ai/blog/karpathy-hackertube/. Try CellCog free, no credit card needed: https://cellcog.ai/signup