HackerTube Review: Karpathy's YouTube With No Algorithm, Tested and Rebuilt in 1:55

You open YouTube to learn one thing, and the feed hands you thumbnails yelling at you and the same five topics you clicked last week. The algorithm isn't built to teach you. It's built to keep you watching.

Andrej Karpathy hit the same wall and had Claude Code build HackerTube: a YouTube with no algorithm, made only from videos Hacker News kept recommending for twenty years. We tested it against YouTube with blind judges, then rebuilt our own version from Karpathy's data with one prompt. It took 1 minute 55 seconds.

Karpathy built HackerTube in two hours

Karpathy co-founded OpenAI and ran AI at Tesla. On a Saturday morning he opened Claude Code and asked, roughly, to build his own YouTube using only videos Hacker News keeps recommending. Two hours later, hackertube.org was live: no login, no clickbait thumbnails, about 6,400 videos real people kept pointing each other to.

Then he gave the data away, with a note in his post: point your Claude at it and build your own. So we did.

The thumbnail trick

YouTube lets creators upload a custom thumbnail, and big channels test dozens of them to find the one you can't help clicking. HackerTube throws that away. It shows a frame from the middle of the video instead, so you see what the video actually is.

No login, Longs and playlists

• No account. Likes are saved in your own browser, and the For You page changes right there. Nothing is sent to anyone.
• Longs. A tab for talks over an hour.
• Playlists. For example the HN canon: the ten videos Hacker News has pointed people to most, year after year.

A fun one: the number one AI video on the whole site is Karpathy's own. Seventeen different people on Hacker News recommended it, years before the site existed. Karpathy didn't rig that. Hacker News did.

43 million comments to 6,454 videos

Here's the funnel, in plain steps:

1. Start with every Hacker News comment ever written: 43 million.
2. Keep only the ones that link a YouTube video and actually say something, at least 100 characters.
3. Keep a video only if three different people mentioned it, in three different months. One fan can't game that.
4. Have Claude read about 40,000 of those comments and label each one: recommending, joking or trashing.

That last step is the clever bit. Only about four in ten mentions were real recommendations. Without the labels, a video everyone hates would look as popular as one everyone loves. The ranking isn't views or likes. It's how many different months people kept saying: watch this.

What Karpathy did by hand

Three things: he made a Google API key, rented a server and bought the domain. Everything else, he just talked to Claude. In his words: "I don't read or write code anymore. I only talk to Claude."

Test 1: HackerTube vs YouTube

Is it actually better, or does it just feel better? We picked five things anyone might want to learn: how large language models work, math for machine learning, how computer chips are made, how to give a good talk and how the economy works. For each, we took the top ten results on YouTube (logged out, no history) and the top ten on HackerTube. That's 100 videos.

We shuffled all 100, hid where each came from, and gave them to two separate Claude judges. One only looked at thumbnails. The other read titles and descriptions and answered one question: would a good teacher send a student to this video?

Honestly, YouTube search was cleaner than we expected: only one shocked face in fifty. The wall of slop Karpathy talked about lives more in the home feed than in search.

The big one: for HackerTube videos on the topic, the teacher judge said yes to 16 out of 17, 94%. On YouTube, 62%.

The catch: a library, not a search engine

HackerTube only has what Hacker News loved, so it won't have a video for every question. When we asked how the economy works, it gave one great pick, Ray Dalio, then wandered into airlines and video-game economies. YouTube found something on topic almost every time.

So think of it this way. YouTube is the search engine. HackerTube is the library of videos people still recommend years later. If you want this week's AI news, use YouTube. If you want to really learn something, start on HackerTube.

Test 2: our own version, one prompt

Under Karpathy's Hacker News post, the top requests were the same: let me sort by newest, let me filter by length, where's the code? The code isn't public. The data is. So that was the test.

We started three brand-new Claude Code sessions with Karpathy's setup: Opus 5.5 on medium effort. Each got the data zip, a YouTube access token and exactly one prompt asking for an AI-only HackerTube with newest-first sorting, a length filter and videos that play inside the page. Then we didn't touch anything.

Each session first opened the instructions file Karpathy wrote for AI agents and read it. It checked the token with one test call, wrote a small script to keep only the AI videos and fetch their dates and lengths from YouTube, then wrote the page, started it and checked its own code.

One prompt, about two minutes, and every site worked: sort by newest, length filter, video playing in the page. At API prices each run cost about 60 cents. Karpathy needed two hours because he built the data from nothing. Starting from his data, it took two minutes.

One thing changed every time: what counts as AI. Two sessions picked 341 videos. The third picked 576, counting data science and robotics, and let in a video about the Game of Life. Same prompt, different judgment.

The clickbait came back

Look closely and all three copies made the same mistake. They used the uploaders' thumbnails, the exact thing HackerTube throws away. Karpathy's instructions even explain the trick. We just didn't ask for it.

The fix was one sentence: "The thumbnails are the uploaders' clickbait ones. Use the frame from the middle of the video instead, like HackerTube does." Sixteen seconds later, it was calm again. The lesson: the AI builds exactly what you ask for, so the part that's still yours is knowing what to ask.

The fair objections

• "It's just a playlist." That was the top pushback. Someone also pointed out a person built almost the same thing a year earlier. Both are fair: the idea isn't new. What's new is one person doing all of it, the 43 million comments, the labelling, the website and the server, in a morning, by talking.
• The labels matter. Someone built the same thing for books within a day, but counted every mention as a recommendation. A book people mostly argue about showed up as recommended by 651 people. Karpathy's version checks whether people liked it. That's the difference between a list and a recommendation.
• The labels aren't perfect. Claude wrote them, and the data file itself warns a few percent will be wrong.

Not a coder? Use it tonight

You don't have to touch code. Just use the site. Start with the HN canon playlist, or "Learn deep learning from zero" if you want to understand AI properly. A feed with no algorithm is also a good gift for a parent who keeps getting pulled into the same videos every night.

The bigger idea: any place where people recommend things to each other can become your own feed. People made a books version and a Wikipedia version within a day.

Get the No-Algorithm Watchlist

We put the 25 AI videos Hacker News recommends most, in the order to watch them, plus our exact prompt, into a free guide. Comment TUBE on the video, or grab it below.

How we tested: Test 1 used the top 10 results for 5 queries on YouTube (logged out) and HackerTube, 100 videos judged blind by two Claude judges. Test 2 used three fresh Claude Code sessions on Opus 5.5, medium effort, one prompt each, hands-off; cost is API-equivalent. Small tests, not a study. Data: HackerTube (hackertube.org), ODC-By, from the open-index/hacker-news dataset. Independent, not affiliated with HackerTube, YouTube or Y Combinator.