Karpathy rebuilt his video feed from 20 years of Hacker News recommendations

43 million comments were the input, not the playlist

The linked dataset lists 43,062,301 comments. Karpathy’s Hacker News submission gives 39,834 recommending comments behind those 6,454 videos. These are three different quantities. According to his write-up, Opus 5.5 filtered the archive and used Sonnet to classify candidates, producing a catalogue across 150 topics. This does not mean a language model individually read every source comment.

Sources: Andrej Karpathy [1] · Andrej Karpathy / Hacker News [2] · OpenIndex / Hugging Face [3]

Even the thumbnail makes a choice

Karpathy says thumbnails use frames from each video’s midpoint, while preferences stay in the browser without an account. He reports roughly two hours of work, including an export for building alternative versions. He personally handled the Google API key, server and domain. The result still depends on human setup and a particular community’s accumulated taste.

Sources: Andrej Karpathy [1]

Choose the filter, not just the next video

Our reading: this makes recommendation design tangible. A neutral-looking thumbnail cannot make a weak lecture useful, and repeated endorsements can reproduce a group’s blind spots. But a catalogue built around an explicit community gives the viewer a clear question to ask: do I trust these people’s reasons for watching? That is a more inspectable basis for discovery than an unexplained sequence of suggestions. The opportunity is to choose whose judgment shapes the queue, while keeping the freedom to change it.