Karpathy builds his own YouTube in two hours – and Claude writes all the code

What's going on

Andrej Karpathy, OpenAI co-founder and former head of AI at Tesla, has built his own little video platform: HackerTube, available at hackertube.org and, according to Cryptobriefing, live since October 31, 2026. The idea: instead of a recommendation algorithm optimized for clicks and watch time, the site only shows videos that users of Hacker News – a well-known discussion forum for programmers, founders and tech enthusiasts – have recommended over and over in their comments.

The result is 6,454 videos sorted into 150 topics. What's most remarkable is how the site came together: by his own account, Karpathy wrote practically none of the code. He had Anthropic's AI model Claude do the work, which took about two hours in total. According to the Ukrainian portal dev.ua, Karpathy – who coined the term "vibe coding", meaning programming by simply telling the AI what you want – joined Anthropic himself this year. That makes HackerTube a showcase for his new employer's product, too.

How the site was built

On X, Karpathy described the project as a weekend toy:

"Weekend hack: HackerTube. I was annoyed by YouTube's recommendations, so I had Claude filter 43 million comments from 20 years of HN down to 6,454 videos that get recommended again and again. Much better viewing experience, ~2 hours with Opus 5.5."

He says he works with Claude Code, Anthropic's coding tool, which runs as a desktop app on the Mac, and uses Opus 5.5 at "medium effort" by default. He just talks to Claude and lets it handle everything. According to a blog post summarized by its author on note.com, the spark came on a Friday evening: YouTube served him overhyped thumbnails and irrelevant recommendations – an example of what the internet calls "enshittification", the gradual decay of big platforms.

According to Cryptobriefing, two models split the work: Claude Opus 5.5 steered the project, while the smaller, cheaper Claude Sonnet did the heavy lifting on filtering. The comments span 2007 to 2026. A video only made it into the collection if it was recommended by multiple users – and in different months. That weeds out short-lived hype while keeping videos with lasting value.

The numbers are a bit fuzzy: Cryptobriefing says the models processed "nearly 40,000 comments". The account on note.com is more plausible: from the full comment pool, around 40,000 posts were identified that actually recommend videos, and the final selection of about 6,500 videos was drawn from those.

Who did what: human vs. AI

By his own account, Karpathy did only three things by hand:

• created an API key in the Google Cloud console (presumably for accessing YouTube data),
• spun up a VPS – a rented virtual server in a data center,
• bought the domain.

Everything else – coding, running terminal commands, deploying the site to the server, security settings and optimization – was handled by Claude. The compute involved was substantial: according to Cryptobriefing, around 504 million input tokens (chunks of text the model reads) went through the system, 90 percent of them from cache – context that was stored and reused, which is much cheaper. On top of that came 7.6 million output tokens. None of the sources say what it cost in euros.

What HackerTube does differently

The site deliberately ditches typical YouTube mechanics:

• No clickbait thumbnails: Instead of the uploader's chosen thumbnail, HackerTube shows a frame from the middle of the video (at 50 percent of its runtime). That gives you a better idea of what it's actually about.
• No account required: The "For you" recommendations work without logging in. Your history is stored only locally in your browser, so the site doesn't need to know who you are.
• Keyboard shortcuts for navigation.
• Full data export: You can download the entire dataset as a ZIP file and process it yourself – for example, by handing it to your own AI agent and having it build a personalized version.

According to the counter-test, the videos include plenty of in-depth computer science lectures, programming language history and clever hardware hacks – exactly what the Hacker News community values. Cryptobriefing rightly points out that the collection reflects the taste of exactly one community, so the 150 topics are correspondingly skewed.

The counter-test: Would simple counting have worked?

Developer Saqib (GitHub user saqibns) asked the obvious question: why do you need an expensive language model to sift through millions of comments? You could just count which YouTube links show up most often. He downloaded a public Hacker News archive from Hugging Face (41.8 million posts and comments from October 2006 to November 2023) and ran a script to count all YouTube links. The run took under five minutes and cost nothing. It found 168,756 YouTube mentions of 129,141 distinct videos.

He then compared his top 600 with 600 videos from HackerTube:

So the counting approach produced an almost entirely different list. The most-linked videos included a nutrition talk about sugar (119 mentions), "Don't Talk to Police" with legal advice (106) and a talk on motivation (82) – videos that get posted reflexively as ammunition in relevant debates. Only at number five did something appear that the tester would expect in HackerTube: Linus Torvalds on Git. Sorted by upvotes, viral news moments dominated, like the first photo of a black hole or Boston Dynamics robots.

Particularly telling: 330 of the 417 HackerTube videos found in the dataset had only 1 to 6 mentions. A frequency filter would have discarded all of them. The tester's conclusion: counting measures volume, Karpathy wanted taste – Claude could read why someone linked a video, not just that they did. One caveat: his archive ends in 2023 and isn't identical to Karpathy's data, which likely explains some of the missing 30 percent.

Analysis

HackerTube isn't a big product, just a weekend project – but a revealing one. It shows how far AI agents have come by fall 2026: a working, publicly accessible website with a curated dataset gets built in an afternoon, while the human just states the goal and clicks through a few web dashboards. That backs up Karpathy's claim that projects that used to take weeks or months can now be done in a morning – though it comes from one of the most experienced AI practitioners in the world, who now works at Anthropic. The project doesn't prove how well this works for non-experts without his know-how.

The division of labor is interesting, too: a large model as project lead and a small model for routine work – a pattern that cuts costs. The counter-test makes the strongest case for using AI here: it wasn't about raw compute but about judging content. Open questions remain: how much of the selection really comes down to Claude's "taste" versus Karpathy's instructions – and the fact that an algorithm, just a different one, still decides what you get to see.