Karpathy had Claude build HackerTube

HackerTube is a website that looks like YouTube. It only shows videos that people on Hacker News, a site where users share and discuss links, have written about since 2007. Andrej Karpathy writes on X that it took about two hours to make the site with Claude Code, Anthropic's AI tool that writes and changes code, the instructions a computer follows, and runs commands. According to the site, the video list is based on comments that Claude, Anthropic's AI model, has read and labeled, and all the data behind the site can be downloaded and used to make your own version.

Why HackerTube exists

Karpathy writes that he was annoyed with YouTube's recommendations. HackerTube gives the same reason. According to the site, YouTube's recommendations are tuned for watch time, meaning they aim to get people to watch as much as possible, so they are very good at finding the next thing you will click, and not so good at finding the talk you will still be thinking about next year.

Meanwhile, the site says, people on Hacker News have spent nearly two decades pointing each other to talks, lecture series and documentaries. Those tips are scattered across millions of comments, and HackerTube collects them in one place.

How the comments became a list of videos

Karpathy writes that Claude filtered down 43 million comments from 20 years of Hacker News. That figure appears only in the post, not on HackerTube. The site describes its own method in five steps.

1. Every Hacker News comment from 2007 to 2026, taken from a public dataset of the full archive, was scanned for YouTube links.
2. A video made the cut if at least three different people had linked to it with a comment of their own, in at least three different months.
3. Claude read each comment along with the title of the story it was posted under, and judged whether the comment recommends the video, cites it as a source, uses it as a joke or criticizes it.
4. Claude wrote a small info box, which the site calls a card, for every video. The card says what the video is about and why people on Hacker News like it.
5. Videos were ranked by how many different months people had recommended them. A video linked again and again over several years counted for more than one that was popular for a short time.

According to the site, the result is 6,454 videos and 39,834 labeled comments, sorted into 150 topics in 16 areas. Not every video is recommended. The explanation file DATA.md, which comes with the data, says about 25 percent were only mentioned or cited and never recommended. The site also says the labels were written by an AI model and that a few percent may be wrong.

In a reply under the post, Karpathy writes that Claude first filtered out a lot by looking at web addresses, using pattern searches and dropping spam. A pattern search looks for fixed patterns in text. Only the remainder was read by the model Sonnet, one of the Claude models.

What you can do on the site

Karpathy thinks the recommendations are excellent ("Excellent recs") and the viewing experience clearly better than on YouTube. The site has these features.

• You get recommendations without logging in. Likes, history and the list of videos you want to watch later are stored only in your browser.
• Each video is shown with a frame from the middle of the video instead of the thumbnail the uploader chose. According to the site, such thumbnails are designed to grab your attention with big faces, arrows and shouting text. Karpathy writes that they are heavily optimized through A/B tests, where two versions are tried to see which one more people click.
• Longs is a separate page with every video over one hour. The site calls it the opposite of Shorts, which are YouTube's short videos.
• The home page mixes classics, hidden gems that few people have watched, recent favorites and topics you have not explored.
• Search understands topics, not just words. According to the site, a search for "distributed systems" finds talks that never use those words.
• Every video page shows what people on Hacker News wrote about the video, with links back to the discussions.

The image in the post shows the home page.

You can use the data to make your own version

Everything behind the site can be downloaded as a single 17.8 MB zip file, one packed file that unpacks into several files. Karpathy writes that you can point your own Claude at the file and ask for your own version. The site suggests the same, using an AI agent such as Claude Code. An AI agent is an AI program that can carry out tasks in several steps on its own. According to the site, it can make a different interface, a weekly email of talks on your topics or an analysis. The explanation file DATA.md is written for such agents and explains how the files fit together.

The data is released under the ODC-By 1.0 license. According to the site, you can use it for anything, as long as you credit HackerTube and the open-index/hacker-news dataset, which the comments come from. The comments still belong to the people who wrote them. YouTube's own information about the videos, such as descriptions, view counts and lengths, is not included. Each video still has its YouTube ID, so you can fetch the information yourself.

What Karpathy did himself, and what Claude did

Karpathy writes that he used Claude Code as a desktop app for Mac, with the model Opus 5.5, also a Claude model. He only talked to Claude and had it do everything, except three things he did by hand in the browser.

1. He created an API key in Google Cloud. An API key is a kind of digital pass that lets a program fetch data from a service, here information about videos from YouTube.
2. He started a VPS, a rented server. A server is a computer in a data center that is always on and shows the website to the people who visit it.
3. He bought the domain name, the web address hackertube.org.

Once the server was up, he gave Claude the IP address, the server's numeric address on the internet. He told Claude to update, secure, harden, deploy and optimize the code. Here, to harden means to make something harder to attack. According to Karpathy, Claude did the whole thing.

The credits on HackerTube say the same. They state that the site was built by Claude Opus 5.5 in Claude Code at Karpathy's request, and that teams of Claude Sonnet 5.5 agents coordinated by Opus wrote the labels, the cards and the topic list.

Karpathy's replies in the thread under the post

A user missed arxiv sanity. Karpathy replied that he had to take it down because it was too much work to maintain, and that he now likes and uses alphaXiv, which he thinks is in the same spirit. alphaXiv describes itself as a place to read, discover and discuss research papers. He answered the question "Open Source?" with a made-up instruction to Claude in his own words, probably as a joke, and not with an answer about whether the code is open.

Karpathy's view of what is now possible

Karpathy thinks that many great projects that would normally have taken weeks or even months can now be made quickly with just a few instructions. He calls it "few-shotted for breakfast." "Few-shot" means getting an AI model to solve a task with only a few instructions or examples, and "for breakfast" means it is quick and easy.

This is his judgment. The time of about two hours and the division of work are also his own account, and we have not been able to check them independently. HackerTube is one example, and one example does not show that the same holds for every project.

Other things you can try to make yourself

Having an AI read and sort large amounts of text, which is what HackerTube is built on, can in principle also be used on other collections of text. Here are two examples you can try yourself.

• A made-up example is collecting the books or podcasts people recommend to each other in comments on another website, the way HackerTube collects videos. First check that the website allows the comments to be collected.
• Another made-up example is going through your own notes or documents to get an overview of books, links or ideas you have saved over the years.

You can start by asking an AI assistant, a program such as Claude or Claude Code where you write in plain words what you want and get an answer or a finished result back. Describe what you want, which texts it should use and what the result should look like. Think about privacy before you share your own documents with an AI service, especially if they contain information about other people. The result may not turn out as well as HackerTube.

Key takeaways

• HackerTube collects YouTube videos that people on Hacker News have written about since 2007. According to the site, about 25 percent of them were only mentioned and never recommended.
• Claude filtered and labeled the comments, and according to the site a few percent of the labels may be wrong.
• The site shows recommendations without logging in, frames from the middle of each video and a separate page for videos over one hour.
• According to the site, the data can be used freely if you credit HackerTube and open-index/hacker-news, and the comments still belong to the people who wrote them.
• That the site took about two hours, and that many projects can now be made quickly, is Karpathy's own account and judgment.

Sources and further reading