RU

Karpathy rebuilt YouTube in an evening with Claude Code

Published: 2026-11-02 · Author: AI Release · @ai_release1
Karpathy rebuilt YouTube in an evening with Claude Code

Karpathy pulled the recommendation algorithm out of YouTube and showed what is left: a list of videos. Out of that list he built a platform. In an evening.

Andrej Karpathy — OpenAI co-founder, former AI director at Tesla, at Anthropic since this year — made Hackertube: a video site running on a YouTube catalog. Claude Code wrote the code. Not a team of twenty engineers, not six months of development. One human and one agent.

Karpathy is not a blogger with a channel about vibe coding. He helped start OpenAI and spent years building Tesla's self-driving stack. When a person like that says he built a platform with an agent, it stops being an experiment and becomes a signal.

What he actually proved

A video-platform clone surprises nobody. What surprises is the split: the agent wrote, the human directed. The industry has promised "describe the task, get the app" for years and tested it on todo lists and landing pages. Here it was tested on a catalog, search and a player.

Then comes the uncomfortable part for platforms. Once you have a catalog and a player, recommendations become an option, not a requirement. Turn the algorithm on if you like. Or just scroll. While YouTube decides what you see next, Hackertube hands the question "what do I want to watch" back to the user.

Both outlets land on the same point: the value is not the site itself, it is that anyone can now build such a site. ifanr puts it bluntly — everyone can have an internet of their own.

Why this matters now

The programming profession is not disappearing. The entry barrier is. Building "your own YouTube" used to mean a team and a budget. Now it means a free evening and a model subscription. Bad news for anyone selling development by the hour. Good news for anyone with something to build.

There is a second layer. Hackertube lives on someone else's catalog. YouTube hands over metadata, not economics: ads, rights, monetization stay outside. A personal internet assembles fast, but it runs on infrastructure somebody else holds and can shut off whenever they want.

That is why this reads less like "look what the model can do" and more like a warning to platforms: your moat is not the code, it is the habit of users typing your name first.

💬 What to argue about

  • A personal catalog instead of recommendations: freedom of choice, or a faster way into your own bubble?
  • If a platform takes an evening to build, what actually protects the product — the code, the content, or the habit?
  • Does agentic coding take work from juniors, or remove the barrier for people who never programmed?

Prediction: Within a year there will be dozens of personal "own internets" built on agents, and big platforms will start selling recommendation access as an API — otherwise they lose the entry point to content.

Disclaimer: not officially confirmed

TechNews

🤖 AI summary
#Karpathy#Claude#YouTube
📖
Read the guide on this topic
Read →
← PreviousNOIR scans Android apps via adb: where data leaks and how to roll back

Source: technews.tw · post in Telegram