Karpathy Lets Claude Build HackerTube From 6,454 Hacker News Videos

Andrej Karpathy has published HackerTube, a YouTube-style site holding exactly 6,454 videos that Hacker News commenters recommended to each other between 2007 and 2026. He says Anthropic's Claude did nearly all of the work, and that he built the entire product in roughly two hours.

Karpathy, an OpenAI co-founder and former head of AI at Tesla, joined Anthropic in May. In a blog post dated Oct. 31, he explained the motive bluntly: YouTube's home page had become "an enshittified wall of slop" for him. "I'm sure this algorithm works well for the median person on the internet, but it definitely does not work for me," he wrote.

The Decoder, which covered the project on Sunday, reported that Karpathy considers the result a "significantly better viewing experience" than YouTube's own recommendations. The site sorts its videos into 150 topics.

How was the site actually built?

The pipeline is the interesting part. Karpathy ran Claude Opus 5.5 on medium effort inside the Claude Code desktop app. Claude downloaded the Hacker News comments and filtered them, and Karpathy then ran the cheaper Sonnet model over the results in a Claude Code Workflow to strip out spam.

That pass left roughly 40,000 comments that recommend a YouTube video. Claude then built the interface and handled deployment, security hardening and optimization on a rented server.

Karpathy's manual contribution was minimal. He did three things by hand:

• created a YouTube API key in Google Cloud to fetch video metadata;
• booted the server and gave Claude its IP address;
• bought the domain.

"I only talk to Claude and have it do everything," he wrote. A project like this "takes ~2 hours now," he added. He did not disclose what the build cost.

The message is that agent-driven development has compressed a task once requiring a small team — data engineering, front-end work, deployment, security — into a conversation.

What does the product do differently?

HackerTube takes a deliberately spare approach. It has no accounts at all. It shows a frame from the middle of each video instead of the uploader's chosen thumbnail, which removes the incentive structure that shapes clickbait imagery on the platform.

"For You" recommendations stay in the browser's local storage rather than on a server, so no profile accumulates anywhere the user does not control. The site also offers a data export, so other people can hand the list to their own AI agents and fork the experience.

In short, the site is a working argument that a curated static corpus can beat an engagement-optimized feed.

Was this a one-off?

No. Karpathy then built a second site, TristanTube, from a list of about 47,000 videos compiled by Tristan Hume. He said that build was "one-shotted with Claude Code in 3 minutes," according to The Decoder.

One caution applies to the numbers. The build times are Karpathy's own accounts, and no one has independently reproduced them. The two-hour figure covers a developer who knows precisely what to ask for, not a novice.

Even so, the pattern is clear: with a capable agent, a filter over a public dataset and a rented server now constitute a complete product plan.