Andrej Karpathy's experiment: building a web platform with Claude in 2 hoursArticle · digest for November 7
Screen showing rapid web platform development in vibe coding style
Article4 sections≈ 6 min

Andrej Karpathy's experiment: building a web platform with Claude in 2 hours

The phrase "vibe coding" has become one of the most talked-about memes among developers in recent months. It refers to a style of programming where a human doesn't write code by hand in a traditional editor at all, but simply chats casually with a neural network over a cup of coffee, phrasing wishes in natural language. Skeptics have tended to brush off the practice, treating it as a toy for amateurs building trivial pet projects.

But when the concept is put to the test by Andrej Karpathy—former AI director at Tesla, OpenAI co-founder, and one of the world's most respected machine learning authorities—the experiment deserves serious attention. Fed up with modern YouTube recommendation algorithms ("an endless wall of clickbait thumbnails and meaningless noise"), Karpathy decided to build his own alternative video search platform, HackerTube, in a single Saturday morning. And he pulled it off in exactly two hours using Claude Opus 5.5 and terminal Claude Code.

Saturday hackathon against recommendation noise

The idea behind HackerTube was ruthlessly pragmatic: bring the internet back to an era of high-quality educational content. Karpathy recalled that on the legendary Hacker News forum, users over the past twenty years had shared links to remarkable, deep lectures on math, physics, programming, and history, accompanied by thoughtful reviews.

Karpathy gave Claude an ambitious task:

  1. Pull and parse a gigantic 20-year dataset: the complete archive of Hacker News comments from 2007 to 2026, totaling exactly 43 million messages.
  2. Filter out spam, short emotional outbursts, and broken links.
  3. Extract genuine community recommendations and compile a structured catalog of the best educational videos.

43 million comments in 15 minutes: the filtering pipeline

A traditional engineering approach to this task would have required writing complex data processing pipelines, tuning regular expressions, and spending hours debugging batches. Under Karpathy's direction, Claude parallelized the filtering scripts: the archive of 43 million comments flew through the pipeline in under 15 minutes.

Out of 40,000 video hosting mentions found, the algorithm stripped out commercial ads and assembled a golden catalog of 6,454 unique videos grouped across 150 subject categories.

Meanwhile, the HackerTube web interface was designed with a principled rejection of clickbait:

  • No flashy thumbnails: the service automatically grabs a random frame from the middle of a video, showing the actual contents of a blackboard or lecturer's screen instead of creators' exaggerated faces.
  • No recommendation feeds or trackers: the site has no user accounts, profiling, or cookie collection—watch state is saved strictly in the browser's local storage.
  • Open data: the entire curated link database is exported as an open JSON file so anyone can spin up their own mirror of the service.

The engineer's role in the vibe coding era: the Director-Worker Pattern

The most revealing part of the experiment is how labor was split between human and neural network. Across the entire two hours, Karpathy personally wrote exactly 0 lines of Python, JavaScript, or HTML code.

His physical involvement came down to just three actions:

  1. Going into the Google Cloud console and generating an API key for the official YouTube Data API v3 to legally fetch video metadata.
  2. Renting a basic virtual private server (VPS) from a hosting provider.
  3. Registering the domain name hackertube.org.

All the rest of the engineering work—the SQLite database schema, Python backend, minimalist frontend, Nginx reverse proxy configs, Let's Encrypt SSL certificate generation, and systemd auto-restart scripts—was written and deployed from start to finish by Claude.

Does that mean the programming profession is dead? Quite the opposite. Karpathy emphasized that the model operates as a brilliant but literal executor (Director-Worker Pattern). If the operator didn't have a fundamental grasp of network protocols, caching principles, and API rate limits, he wouldn't have been able to catch and correct Claude's mistakes when configuring CORS headers or hitting daily request quotas.

Service architecture and recommendation extraction script

Here is the core link-parsing and filtering module that Claude generated to Karpathy's specification:

import re
import json
from typing import Dict, Any, List

YOUTUBE_URL_PATTERN = re.compile(
    r'(https?://(?:www\.)?(?:youtube\.com/watch\?v=|youtu\.be/)([a-zA-Z0-9_-]{11}))'
)

def extract_hn_recommendations(comments_pool: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
    curated_items = []
    
    for entry in comments_pool:
        comment_text = entry.get("text", "")
        match = YOUTUBE_URL_PATTERN.search(comment_text)
        if not match:
            continue
            
        full_url, video_id = match.group(1), match.group(2)
        
        # Spam-filtering heuristic: comment length at least 60 chars, no stop words
        if len(comment_text) < 60 or "subscribe" in comment_text.lower():
            continue
            
        curated_items.append({
            "video_id": video_id,
            "canonical_url": full_url,
            "hn_discussion": comment_text[:280].strip(),
            "community_score": entry.get("score", 1),
            "created_utc": entry.get("time")
        })
        
    return curated_items

And to deploy the web server on the target VPS, a generated automation script was used:

#!/usr/bin/env bash
# Automated setup of the HackerTube environment on Ubuntu 24.04
set -euo pipefail

apt-get update && apt-get install -y python3-venv nginx certbot python3-certbot-nginx
python3 -m venv /opt/hackertube/venv
/opt/hackertube/venv/bin/pip install fastapi uvicorn httpx
systemctl enable --now nginx
echo "Service infrastructure is ready for launch."

Karpathy's experiment vividly outlines the new reality: the barrier between an abstract idea in your head and a live global product has shrunk to a single Saturday morning. The winners aren't those who type fastest, but those who think clearly, formulate architectural contracts, and steer the power of models in a creative direction.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 3