Fed up with YouTube’s “garbage wall”—Karpathy gets Claude to produce a clean version in two hours.
Fed up with YouTube’s “trash wall”! Karpathy gets Claude to churn out a clean version in two hours
Content Summary
Xinzhiyuan reports Friday night at 11 PM, Karpathy was tired, he opened YouTube, just wanted to find something to watch casually. He was greeted by an entire “trash wall”: each thumbnail more exaggerated than the last, recommendations looping over and over
Xinzhiyuan Report
Xinzhiyuan Report
At 11 p.m. on Friday night, Karpathy was tired, opened YouTube, and just wanted to find something to watch.
What greeted him was an entire “wall of trash”: each thumbnail more exaggerated than the last, and the recommendations kept looping over the same old topics.
He complained: the whole screen is filled with content at the level of “Dumb Evolution Theory”.
After a long day and wanting to relax, I only ended up scrolling through a bunch of clickbait: you’re definitely not unfamiliar with this feeling!
In Karpathy’s view, YouTube’s recommendation algorithm may please the majority, but it fails to satisfy users like him.
Later he said in the comments that users like him are too few: in the eyes of the algorithm, they’re so small they can be rounded off.
This time, he decided he couldn’t tolerate it.
The next morning, he had Claude build a YouTube that only hosts good videos. The site went live that same day.
Karpathy named the site HackerTube.
6,454 videos and 150 topics make up Karpathy’s own video-watching portal.
The videos are still from YouTube, and the recommendation cues come from Hacker News (HN), a tech forum crowded with programmers and entrepreneurs.
Karpathy’s idea is simple: Collect the videos recommended by HN users into playlists.
He handed almost the entire task over to AI.
Claude first downloaded 43 million comments from HN spanning nearly 20 years, filtered out the junk, and kept only those recommending YouTube videos—about 40 thousand.
In the end, 6 454 videos were organized into 150 categories.
The titles and durations were pulled from YouTube, then wrapped in a YouTube‑like interface and deployed.
The whole project took about two hours.
From data screening to deployment, Karpathy says he neither read nor wrote code.
He also packaged and released the data publicly.
Anyone can feed it to their own AI and create their own version.
On Saturday morning, he gave Claude a command:
Following the video recommended by HN users, make our own YouTube.
After that, Claude basically did the work.
Opus 5.5 handled coordination, while the task of filtering comments was given to the more agile Sonnet to batch process.
After stripping out the junk, what remains are the comments on the recommended videos.
The website is finished, but it still needs to go live.
Karpathy spins up a server and sends the IP address to Claude.
Updates, deployment, security hardening, and performance tuning—Claude handles it all in one go.
Meanwhile, this OpenAI co‑founder only does a few errands like setting up the server and buying the domain during Claude’s downtime.
He also joked that it felt a bit like teamwork.
He joined Anthropic’s pretraining team in May this year.
This time he used his own Claude.
The most striking part is the line Karpathy added at the end of the blog:
He no longer does any traditional engineering work.
He doesn’t read or write code, doesn’t type commands in the terminal, just talks to Claude.
A netizen pointed it out directly in the comments: this sentence is the main point of the whole text.
Someone who once led Tesla’s AI team now works in only one way: making requests.
In 2021, Karpathy (left) spoke at Tesla AI Day, with Musk on the right.
HackerTube site looks just like YouTube at first glance, but using it is the complete opposite.
It doesn’t require login; what you watch only exists in your own browser.
There are no repeatedly tested, clickbait thumbnails.
Each video’s thumbnail has been replaced with a frame from the middle of the video. There’s also a “Longs” entry in the sidebar, dedicated to long videos.
The name obviously targets YouTube Shorts.
Karpathy posted the site on HN, and the thread surged past 2,000 points. The comment section accumulated over 500 replies.
Netizens compared it with YouTube’s homepage.
Someone said the homepage recommendations are even more accurate than YouTube’s personalized recommendations.
Someone dived into the rabbit hole of Bach work analyses and said you can’t find them on YouTube.
Some people say that switching from YouTube feels like walking into a forest.
A netizen said his 87‑year‑old father watches YouTube every day, constantly being fed negative World War II content, and can’t get out of it.
Even geriatric doctors are urging seniors to stop watching these kinds of videos.
This netizen believes that HackerTube is already very close to what fathers need.
The more you glance at something, the more the algorithm pushes of it. Many people, like the elderly person above, are trapped in an information echo chamber.
Some netizens complain that the algorithm cares about how to keep you watching; others say the platform doesn’t care about how you feel after watching.
Even paying users are fuming.
A netizen said they have been a YouTube Premium subscriber for several years, but the more they use it, the more they hate it.
Offline sync often fails, making it impossible to watch their own content on a plane.
The netizen lamented, “I really wish I could build my own YouTube client, but I’m not sure if the platform would allow it—so frustrating!”
A website built in two hours hit a large group of people’s resentment toward algorithms precisely.
Some netizens see even further ahead.
This netizen believes that in the past, making recommendations and building indexes were very difficult.
Now, we can probably create our own recommendation system according to our own tastes.
In his view, platforms are a group of “eye‑selling agents” who only want you to keep watching, and the content therefore becomes increasingly extreme.
Now is the time to get rid of them.
People do less manual work, but the backend work hasn’t decreased.
Someone on HN asked how many tokens this project actually burned and how much it cost.
Karpathy had Claude do the math itself.
Approximately 504 million tokens were input, with about 90% hitting the cache.
Outputs about 7.6 million tokens.
More than 90% of the usage is spent on Sonnet tagging comments.
When actually building and deploying the website, Opus only used about 1 million tokens.
This 500 million is accumulated from repeated calls. The model does not ingest 500 million tokens in one go.
Karpathy revealed that over 90% of model usage goes to comment annotation, while application development and deployment consume about 1 million Opus tokens.
In a two‑hour window, the most token‑intensive step was tagging comments. Writing the website turned out to be a minor task.
This calculation serves as a reminder to anyone building data tools: cleaning and annotating data may consume more tokens than writing code.
It might be wise to start with a small batch of data for testing; if it works well, scale up.
Karpathy released the comment analysis data as a zip archive.
His advice is: hand it over to your AI and build your own version.
The package also includes a guide written for the AI, explaining how to use these files.
YouTube video information isn’t provided directly in the data bundle; you have to fetch it from YouTube yourself. The guide explains how to retrieve it, typically by applying for a YouTube API key like Karpathy did.
Once the data and the guide are ready, you can fully “personalize” it according to your preferences.
For example, if you only care about programming, you can have the AI keep only the programming lessons. Then have it categorize them by topic, attach the original recommender to each video, and tick off the ones you’ve watched.
When you open it, it’s like a personal schedule. What you want to learn is instantly clear, and you know exactly where you are in your learning.
The same approach can be applied to the resources you have, such as dusty articles in your bookmarks or a tool list you’ve been accumulating for years.
Which videos to watch used to be decided by the platform.
Now, someone who refuses to settle can, like Karpathy, build a personal video site.
In the past, if you wanted a handy tool, you had only two options: wait for the platform to change, or write the code yourself.
Now there’s an extra step: clearly state the requirements to AI and let it handle the work.
A commenter said that the barrier to creating software is lowering. People can quickly build alternatives and are no longer tied to software giants.
Those who were rounded off by algorithms are now creating their own algorithms.
Of course, Karpathy himself is a top AI researcher. An ordinary person following his steps may not be able to finish it in two hours.
But the door has already been cracked open a little.
Karpathy only set one rule: watch videos that HN users have recommended.
The rest is left to Claude.
When it’s your turn to act, first consider three things: where the data comes from, what criteria you’ll use to filter it, and what you ultimately want to achieve.
AI can write code for you, but it can’t set the standards.
Reference:
Edited by: Yuan Yu
Get MirrorSpeed
Get MirrorSpeed