Karpathy builds a video platform over a weekend with Claude

Andrej Karpathy, co-founder of OpenAI and now researcher at Anthropic, built HackerTube in a weekend using Claude Opus 4.5. The model filtered 43 million comments from 20 years of Hacker News and distilled 6,454 videos that continue to be recommended there. The result, according to Karpathy on X, is a "significantly better viewing experience" than YouTube's own recommendations.

In this article you will see:

• What Karpathy built and how he used Claude
• What this says about development pace with coding agents
• The limits of the case and what not to copy
• How to apply the same logic to your project

What was built

The HackerTube does not require login, shows a single frame in the middle of the video instead of optimized thumbnails and stores state locally in the browser. Shortly after, Karpathy built TristanTube, "made with Claude Code in 3 minutes", with 47,000 videos taken from a list of YouTube channels by Tristan Hume.

Karpathy says he let Claude Code handle almost all the work, from data analysis to server setup. His only manual steps were a Google Cloud API key, a rented server and buying the domain. Projects that used to take weeks can be done in a few hours, according to him.

Technical sheet of the case

• Project: HackerTube, video recommendation platform from Hacker News
• Author: Andrej Karpathy
• Model used: Claude Opus 4.5, via Claude Code
• Data processed: 43 million comments over 20 years of Hacker News
• Result: 6,454 distilled videos
• Derived project: TristanTube, with 47,000 videos
• Build time: one weekend

The data volume and clear scope are what make the case viable. A coding agent does not create data from nothing: it organizes what already exists.

What this says about development pace

The case is an extreme example, but it illustrates a real shift. Coding agents like Claude Code can go from draft to published product in hours, when the task is well defined and the data already exists. The speed does not come from the model alone: it comes from the combination of a clear problem, API access and a developer who knows what they want.

It also shows the role of the human. Karpathy defined the goal, provided the credentials and validated the result. The agent executed. Without direction, the same model would produce something generic.

The limits of the case

Not every project fits in a weekend. HackerTube works because the data source (Hacker News) is public and structured, and because the scope was clear: filter, summarize and recommend. Projects with ambiguous requirements, fragile integrations or sensitive data do not follow the same pace.

There is also the risk of confusing speed with quality. A fast prototype can accumulate technical debt if it goes to production without review. Karpathy's case is an experiment, not an engineering standard.

How to apply the same logic to your project

Who wants to replicate the pace can follow these steps:

1. Pick a small problem with accessible data.
2. Give the agent a clear, verifiable goal.
3. Keep credentials and infrastructure under your control.
4. Review the output before publishing.

Those already using code agents in production can compare models on the same project, putting each one in an isolated space and tracking results in one place. The Cult.fit followed a similar path by standardizing operations on Codex and building internal apps, and the Codex 0.167.0 showed how to centralize agents in the terminal.

The practical lesson is simple: start small, measure, and only then expand scope. Code agents do not replace judgment, but they accelerate those who already know what they want.

It is also worth noting that HackerTube does not use its own model. Karpathy chose Claude Opus 4.5 because it solved the problem, not out of brand loyalty. Switching models when the case calls for it is part of the job.