Key points
- Karpathy, described by The Decoder as an Anthropic researcher, made HackerTube in one weekend using Claude's Opus 5.5 model, with Claude Code doing nearly all of the work.
- Claude sifted 43 million Hacker News comments from two decades and surfaced 6,454 videos that users repeatedly recommend there.
- He then spun up TristanTube, covering 47,000 videos from a channel list by Tristan Hume, reportedly in about 3 minutes.
- Human effort stayed minimal: Karpathy set up a Google Cloud key, rented a server and bought the domain, nothing more.
Why it matters for marketers: Small teams can now prototype curated content products from community data in hours, so the edge shifts from building tools to choosing good sources and taste.
Andrej Karpathy, who The Decoder describes as an OpenAI co-founder and current Anthropic researcher, built a video discovery site called HackerTube over a single weekend. He used Claude's Opus 5.5 model to work through 43 million Hacker News comments spanning 20 years, and it pulled out 6,454 videos that commenters keep pointing to.
Karpathy wrote on X that the result gives a "significantly better viewing experience" than YouTube's own recommendations. The site asks for no login, displays one frame from the middle of each video in place of a polished thumbnail, and keeps user state in the browser instead of on a server.
He followed up with TristanTube, which he said Claude Code produced in one shot in about 3 minutes. It covers 47,000 videos taken from a channel list by Tristan Hume. Karpathy says Claude Code handled almost everything, from analyzing the data to configuring the server, and that he only handled account and hosting basics himself.
For marketers and content teams, the notable point is the method rather than the site. Community signals, such as what experienced people repeatedly recommend, can be mined and packaged into a curated product quickly and cheaply. Karpathy says work that once took weeks can now take a few hours.
Worth watching: whether teams and agencies start building similar niche curation tools for their own audiences, and how platform recommendation systems compare with human-endorsed lists. These are one researcher's reported results, so the quality claims have not been independently tested.
Sources
This brief is our own summary. Read the full reporting at the source.