Karpathy built HackerTube with Claude in two hours, featuring only videos recommended on Hacker News

At 11 p.m. on Friday, Karpathy was tired. He opened YouTube, intending to find something random to watch.

What greeted him was a whole wall of "garbage": the thumbnails were more and more exaggerated, and the recommendations kept cycling through the same few old topics.

He complained: the whole screen was full of trash content at the level of Idiocracy.

After a tiring day, wanting to relax, all he got was a bunch of clickbait: you must be familiar with this feeling!

In Karpathy's view, YouTube's recommendation algorithm may be able to please most people, but it cannot satisfy users like him.

Later, in the comments section, he said that users like him are too few: in the eyes of the algorithm, they are small enough to be rounded off.

This time, he decided not to put up with it anymore.

The next morning, he had Claude build a YouTube that only contains good videos. That same day, the website went live.

Karpathy named the website HackerTube.

6,454 videos and 150 topics make up Karpathy's own video-watching portal.

The videos still come from YouTube, while the recommendation clues come from Hacker News (HN), a tech forum where programmers and entrepreneurs gather.

Karpathy's idea is simple: gather videos recommended by HN users into a playlist.

Claude first downloaded 43 million comments from HN over the past nearly 20 years, then filtered out the junk and kept only those recommending YouTube videos, nearly 40,000 comments.

Finally, it organized 6,454 videos and sorted them into 150 topics.

The video titles and durations were then pulled from YouTube. Then a YouTube-like interface was applied and the site was deployed.

The entire project took about 2 hours.

From filtering the data to going live, Karpathy said he neither read nor wrote code.

Following the videos recommended by HN users, build our own YouTube.

After that, Claude basically did the work.

Opus 5.5 handled overall coordination, and the task of filtering comments was assigned to the lighter, faster Sonnet for batch processing.

The junk content was washed out, and what remained were comments recommending videos.

Once the website was written, it still had to go live.

Karpathy set up the server and handed the IP address to Claude.

Updates, deployment, security hardening, and performance optimization were all handled by Claude in one go.

And this OpenAI co-founder only did a few errands such as setting up the server and buying the domain in the gaps while Claude was working.

No reading or writing code, no typing commands in the terminal, only talking to Claude.

One netizen pointed it out directly in the comments: this sentence is the key point of the whole article.

In 2021, Karpathy (left) speaks at Tesla AI Day, with Musk on the right.

It looks like YouTube

but does everything the opposite way

At first glance, the HackerTube website interface looks like YouTube, but using it feels the opposite in every way.

There are none of those repeatedly tested thumbnails designed to lure clicks.

Each video's thumbnail is replaced with a frame from the middle of the video. There is also a "Longs" entry in the sidebar, dedicated to long videos.

The name is clearly aimed at YouTube's Shorts.

Karpathy posted the website to HN, and the post surged to more than 2,000 points. The comments section followed with more than 500 replies.

Netizens compared it with the YouTube homepage.

Some said the homepage recommendations are even more accurate than YouTube's personalized recommendations.

Some fell down a rabbit hole of interpretations of Bach's works, saying they could not find such content on YouTube.

Others said that switching over from YouTube felt like walking into a forest.

One netizen said that his 87-year-old father watches YouTube every day and, because he is constantly pushed negative content about World War II, is trapped in it and cannot get out.

Even geriatricians advised the elderly man to stop watching such videos.

The netizen believes HackerTube is already very close to what his father needs.

The more you glance at something, the more of it the algorithm pushes. Many people, like the elderly man above, are trapped in information cocoons.

Some netizens complained that the algorithm cares about how to keep you watching; others said the platform does not care how you feel after watching.

Even paying users are seething.

One netizen said he had subscribed to YouTube Premium for several years but disliked it more and more.

Offline sync often has problems, and on planes he simply cannot watch his own content.

The netizen sighed that he really wished he could build his own YouTube client, but was not sure whether the platform would allow it. "It's really frustrating!"

In his view, the platform is a group of "eyeball dealers" who only want you to keep watching, and as a result the content becomes more and more extreme.

Someone on HN asked how many tokens this project actually burned and how much money it cost.

To actually write the website and deploy it, Opus used only about 1 million tokens.

That 500 million was accumulated through repeated calls. The model did not read 500 million tokens in one go.

Karpathy disclosed that more than 90% of model usage was for comment labeling, while application development and deployment used about 1 million Opus tokens.

During the two hours, the step that consumed the most tokens was tagging comments. Writing the website was actually the smaller part.

This calculation is a reminder for people who want to build data tools: cleaning and labeling materials may consume more tokens than writing code.

Karpathy made the comment analysis data public as a compressed package.

The YouTube video information is not directly provided in the data package; you have to pull it from YouTube yourself. The instructions explain how to get it, and generally you need to apply for a YouTube API key first, as Karpathy did.

Once the data and instructions are ready, you can fully "customize" it according to your personal preferences.

Open it this way, and it is a private syllabus. What you want to learn is clear at a glance, and how far you have gotten is also clear.

The same approach can also be applied to materials you have on hand, such as articles gathering dust in your bookmarks or a tool list you have accumulated over the years.

What videos to watch used to be decided by the platform.

Now, a person unwilling to settle can also, like Karpathy, build a personal video site.

In the past, if you wanted a handy tool, there were only two paths: wait for the platform to change, or write code yourself.

A netizen in the comments said that the threshold for making software is dropping. People can quickly build alternatives and are no longer bound by software giants.

People who were rounded off by the algorithm have started building their own algorithms.

Of course, Karpathy himself is a top AI researcher. If ordinary people follow suit, they may not necessarily finish it in two hours.

But this door has already opened a crack.

Karpathy set only one criterion: watch videos recommended by HN users.

Everything else was left to Claude.

When it is your turn to act, also think clearly about three things first: where the materials come from, by what criteria to filter them, and what you ultimately want to get.

References:

This article is from the WeChat public account "Xinzhiyuan" (ID: AI_era), author: ASI Revelation, editor: Yuanyu, published by 36Kr with authorization.