Karpathy builds HackerTube from 43 million Hacker News comments

• HackerTube turns 43 million Hacker News comments into a collection of 6,454 YouTube videos. AI researcher Andrej Karpathy says in his announcement that the build took about two hours.
• URL extraction, regexes and spam filters narrowed the input before Sonnet processed the remainder, according to his clarification.
• The site requires no login, keeps state in the browser and offers the recommendation data as a ZIP for reuse, Karpathy says.

Karpathy’s build log starts with a Friday-night encounter with YouTube’s “wall of slop.” HackerTube adds keyboard shortcuts and a “Longs” view; a reader in the Hacker News discussion has already made a books version.

The 43 million figure describes the starting archive. Karpathy’s build log and subsequent clarification describe this pipeline:

1. Download Hacker News comments spanning 2007 through 2026.
2. Narrow candidates with URL extraction, regexes and spam filtering.
3. Send the remainder to Sonnet through a Claude Code Workflow, which he says used subscription tokens.
4. Retain roughly 40,000 comments recommending 6,454 videos across 150 topics.

The prefilter is the most interesting engineering detail in the build: the model workload was the reduced candidate set.

Karpathy lists three choices in the build log that change the browsing experience:

• Thumbnails: frames from the 50% point of each video replace creator-uploaded thumbnails.
• Recommendations: a custom “ForYou” feed draws from the HN-derived collection.
• Accounts: no user accounts; state stays in the local browser session.

The recommendations reflect what HN commenters recommended over two decades. Karpathy’s claim of a “significantly better viewing experience” is his assessment of that collection.

Karpathy used Claude Code’s Mac desktop app with Opus 5.5 at Medium effort, according to his account. He says he only talked to Claude, leaving code and command execution to the agent.

His manual work consisted of three browser-side tasks:

1. Create a Google Cloud API key for querying YouTube video metadata.
2. Provision a VPS to serve the site.
3. Buy the domain name.

He handled those tasks while Claude worked, then supplied the server’s IP address. He says Claude completed deployment, hardening and optimization.

Karpathy offers the underlying comment-analysis data as one downloadable ZIP, explicitly inviting agent-built variants. The shared artifact carries the accumulated recommendation signal, rather than just the browsing interface.

The reactions split between personal software and the remaining barriers to copying it:

• Personal software: Being able to reshape software around individual preferences was what WesRoth found exciting about the project.
• Copyability: Yuchenj_UW joked that “Don’t make any mistakes” was the underrated moat, quoting Karpathy’s playful description of his prompt. In a later post, he named taste and judgment among the human traits that remain moats.

Karpathy has also described the upkeep burden of an earlier project: he said he took it down because it was “just too much work to maintain.” He added that it would have been easier today.