HackerTube
Andrej Karpathy built a complete YouTube alternative in 2 hours using Claude Opus 5.5—without writing a single line of code—demonstrating what post-code software development actually looks like in practice.
Read Original Summary used for search
TLDR
• Karpathy no longer writes code or runs terminal commands—he only talks to Claude and directs it to build, deploy, secure, and optimize everything
• HackerTube curates 6,454 YouTube videos from 40K Hacker News comments (2007-2026), using 50% frame thumbnails instead of clickbait, with local-session recommendations
• Claude handled the entire stack: downloaded HN data, filtered with Sonnet, built the UI, deployed to VPS, secured/hardened/optimized the server
• The only manual work: creating a Google API key, booting a VPS box, and buying the domain name
• This is a concrete example of AI-native development from someone who helped build GPT and Tesla Autopilot—the shift from implementation to taste and direction
In Detail
Karpathy got fed up with YouTube's "enshittified wall of slop"—exaggerated thumbnails and algorithm-driven mediocrity optimized for median users—and decided to build his own version. Using Claude Opus 5.5, he created HackerTube in about 2 hours on a Saturday morning. The site curates 6,454 YouTube videos recommended across 40K Hacker News comments from 2007-2026, organized into 150 topics. It uses the 50% frame as thumbnails (avoiding clickbait), provides custom recommendations without user accounts (storing state in browser sessions), and includes keyboard shortcuts and data export.
What's remarkable isn't the product—it's the process. Karpathy describes his current workflow: "I don't actually do anything resembling traditional engineering anymore (read/write code, run commands in a terminal, or etc.), I only talk to Claude and have it do everything." Claude downloaded all HN comments, filtered them, ran Sonnet over them to remove spam, built the UI, deployed to a VPS, and handled security hardening and optimization. The only manual steps were creating a Google API key, booting a VPS, and buying the domain.
This is a glimpse into post-code development from someone who helped build GPT at OpenAI and led AI at Tesla. When Karpathy says he's stopped writing code entirely, it signals a fundamental shift: the bottleneck in software development is no longer implementation—it's taste, problem definition, and knowing what to build. The "engineering" becomes conversational direction to AI agents that handle all technical execution.