Karpathy didn't ask YouTube to fix its recommendations. He had Claude build him a replacement in a weekend, and the thing that made it work wasn't the model, it was twenty years of human taste.
In an October 31 blog post, Andrej Karpathy described building HackerTube, a video site assembled entirely from links Hacker News commenters have recommended since 2007. He called YouTube's homepage an "enshittified wall of slop." The algorithm might work fine for the median user, he said, but it doesn't work for him. So he skipped the fix and built an alternative, using Claude to do essentially all of it.
The scale of the input is the whole story. Karpathy had Claude pull roughly 43 million Hacker News comments spanning 2007 to 2026 and hunt for links to YouTube videos that kept getting recommended across threads, year after year, by different people with no coordination between them. Opus 5.5 built the extraction pipeline. Sonnet did the filtering work, cutting that pile down to 6,454 videos sorted into about 150 topics. What's left is not what the platform's engagement engine would surface. It's whatever a community of engineers and builders kept pointing each other toward, for two decades, because it was actually good.
Karpathy says the whole thing took about two hours of his own time. He created an API key through the Google Cloud console, spun up a VPS, bought the domain. Everything else, he handed to Claude: update the server, harden it, deploy the site, optimize the code. That division of labor is the real artifact here, more than the video site itself. A person with deep infrastructure experience and a frontier coding model turned a personal annoyance into a working product before Monday.
The details he built in are the tell that this was made by someone who actually uses the product, not a demo built to impress. Thumbnails come from the midpoint of each video rather than a thumbnail baited for clicks. There's no login. Session state, meaning what you've watched and what gets recommended next, lives in the browser rather than on a server tracking you across visits. And Karpathy exported the full dataset, so anyone can pull the same 6,454 videos and rebuild their own version with their own ranking logic on top.
Andrej Karpathy says the AI capability gap is now a sharp funnel Karpathy's October 11 post puts internal frontier-model access at roughly 5,000 people worldwide, versus 6 billion who've barely used an LLM, a gap he says is now a funnel, not a two-sided conversation. - andrej karpathy ai capability gap widens - ai understanding divide between insiders and masses
HackerTube replaces YouTube's engagement-ranked homepage with a catalog drawn from Hacker News recommendations, while keeping viewing history and personalized recommendations in the user's browser.
The claim embedded in the project is sharper than "AI can build a website fast." It's that the scarce resource in recommendation isn't model capability, compute, or even a bigger dataset. It's curation signal, the kind that accumulates slowly when real people vouch for something to other real people over years, with nothing to gain from the vouching. Karpathy is one of the most credentialed people alive on the model-training side of AI, a former OpenAI founding team member and Tesla's former Autopilot AI lead who joined Anthropic earlier this year. For him to spend a weekend concluding that a planetary-scale recommendation system loses to 6,454 links picked by strangers on a forum is a pointed argument about where the actual bottleneck sits in consumer AI right now. It isn't model quality. It's whose judgment the system is built to serve.
Also read: AI Data Centers Are Draining the RAM Supply and Your Next Laptop Pays For It • How to Negotiate an AI Agent Training Data Indemnity Clause • Google tells Gmail users two minutes of clicking is their job now
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.