Andrej Karpathy built a working video site in one weekend. Claude, the AI model Anthropic sells, did nearly all the work. The whole job took about two hours1.
Karpathy once helped start OpenAI. He ran Tesla's AI team for five years after that. Today he runs a small company called Eureka Labs7.
He named the new site HackerTube. It pulls videos from 43 million Hacker News comments posted since 20071. Karpathy says the result beats YouTube's own picks, for him.
The Two-Hour Build
Karpathy laid out the process himself, in a post on X dated Nov. 11. He said he barely touches a terminal anymore. Talking to Claude now does the whole job, he says.
Andrej Karpathy@karpathyWeekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:https://t.co/xULTGY6VhH
- Excellent… pic.twitter.com/QRq80bJo4o
Claude Opus 5.5 read every one of those 43 million comments. It pulled out 6,454 videos that Hacker News users kept recommending across nearly two decades1. Then it sorted the videos into about 150 topics, so a visitor can browse by subject instead of scrolling one long list.
Karpathy still did three small jobs by hand. He requested an API key from Google's cloud console, rented a server, and bought the domain name1.
Everything past that point ran through the agent alone. "Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing," Karpathy wrote1.
The Weekend Copy Machine
Karpathy's idea spread within a day. A coder who posts under the name roansh saw HackerTube and built a matching site for books3.
That new site, HackerBooks, catalogs 16,543 titles that at least two different Hacker News commenters named since 20073. Roansh wrote that Claude built almost the entire thing, data pipeline included, while the human half mostly steered and bought a domain.
Both sites now sit near the top of Hacker News's own front page. Karpathy's thread holds 2,614 points and 577 comments2. HackerBooks, posted a day later, already holds 1,452 points of its own3.
Karpathy's Build Draws 80 Percent More Points Than Its Copy
Points on each weekend AI build's own Hacker News discussion thread, as of Nov. 2, 2026
Each column is one project's own Hacker News thread; Karpathy's HackerTube stands taller than the HackerBooks copy posted a day later.
HackerBooks posted a day after HackerTube, so its thread has had less time to collect points.
Sources: Hacker News [2]; Hacker News [3]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| HackerTube | 2,614 |
| HackerBooks | 1,452 |
Hacker News itself is nineteen years old. Paul Graham and his partners at Y Combinator launched the forum in 20078. Two decades of its comments just became raw material for two fresh websites, built inside one weekend.
The Case for Agentic Speed
Karpathy makes a claim bigger than one video site. "Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast," he wrote1.
He says he quit coding the old way entirely. "I only talk to Claude and have it do everything," Karpathy wrote, describing how he now works1.
Numbers back that claim, beyond a feeling. His post pulled 41,600 likes and 1,716 replies inside one day1. Karpathy ranks among AI research's most watched voices. His demo reached other builders fast, and at least one of them copied it by morning.
The Case Against
Simon Willison builds software for a living and writes often about AI coding tools. He draws a sharp line between a weekend hack and a system other people depend on5.
"Think about how much harm the code you are writing could cause if it has bugs or security vulnerabilities," Willison wrote5. HackerTube serves public video links, skips logins entirely and moves zero payments, a lower-stakes job than most business software carries.
A fresh benchmark backs that caution with harder numbers. SemiAnalysis released GDPx this week, built from real tasks in finance, law and chip design4. The best model solved just 12 percent of them.
Best AI Model Clears Just 12 Percent of a New Economic Benchmark
Top score on SemiAnalysis's GDPx benchmark across finance, law, chip design and other high-value work, released Nov. 2, 2026
The colored wedge is the share of GDPx's tasks the best-scoring model solved; the rest of the circle marks tasks that model missed. Higher is better, and today's top score covers a thin slice of the circle.
SemiAnalysis built GDPx this month, so this first score stands alone until a later run shows whether it rises or falls.
Source: SemiAnalysis [4]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Tasks the best model solves | 12 |
| Whole | 100 |
GDPx's top score came from elsewhere, past coding entirely. On GDPx's own software engineering tasks, Google's Gemini 4 Carbon edged past Claude, the model behind HackerTube4.
What Both Sides Would Need to Be True
Karpathy's case holds if HackerTube's speed turns ordinary, a routine result instead of a highlight reel. Many builders would need to ship working products this fast, on stakes that matter as much as a video feed does.
Willison's case holds if speed and safety keep trading off against each other. GDPx's 12 percent score says most real jobs still beat today's best models by a wide margin4.
Karpathy himself already drew this exact line, months before HackerTube shipped. He calls fast, casual building "vibe coding," fine for a weekend toy6.
The harder discipline, Karpathy calls "agentic engineering," saying it "preserves the quality bar of professional software." "You are not allowed to introduce vulnerabilities because of vibe coding," he wrote. "You are still responsible for your software, just as before."
HackerTube is vibe coding, by its own builder's name for the category. The floor for a weekend toy has risen fast this year. Whether the ceiling for serious, reviewed software moved with it stays a separate question, and this build leaves it alone.
What to do
Judge any demo by its stakes. A weekend site that welcomes any visitor, with real money kept out of the loop, can run on vibes and still turn out fine.
A tool that touches customer data, payments or safety needs the discipline Karpathy himself names: agentic engineering. It calls for review, tests and a person who can explain every change.
By the numbers
- 43 million: Hacker News comments, posted since 2007, that Claude searched to build HackerTube1.
- 6,454: videos HackerTube sorted into roughly 150 topics, all originally recommended by Hacker News users1.
- About two hours: the build time Karpathy reports, running mostly through Claude Code with Opus 5.51.
- 41,600 likes and 1,716 replies landed on Karpathy's post about HackerTube within a day1.
- 2,614 points and 577 comments mark the Hacker News discussion of HackerTube itself2.
- 16,543 books fill HackerBooks, the copycat site a second coder shipped one day later3.
- 12 percent: the share of real-economy tasks the best AI model solved on SemiAnalysis's new GDPx benchmark4.
What to watch
Watch whether HackerTube and HackerBooks still draw visits a month from now. The novelty will fade, and Hacker News will move to its next story. A site that keeps earning return visits says the underlying idea has lasting value, past the weekend it launched in.
GDPx's scores deserve watching too. SemiAnalysis built the benchmark to track how fast that 12 percent climbs as labs release new models. GDPx's climb is the clearest sign of whether agentic engineering is catching up to vibe coding's new speed.
Sources
- Andrej Karpathy, "Weekend hack: HackerTube," Andrej Karpathy on X, Nov. 1, 2026, https://x.com/karpathy/status/2116957135661945283
- "HackerTube: YouTube built from videos recommended by HN," Hacker News, Nov. 1, 2026, https://news.ycombinator.com/item?id=50319510
- roansh, "Show HN: HackerBooks – HackerTube, but for books mentioned on HN," Hacker News, Nov. 1, 2026, https://news.ycombinator.com/item?id=50326821
- Max Kan, Andrew Megalaa and Dylan Patel, "GDPx: measuring frontier model capabilities," SemiAnalysis, Nov. 2, 2026, https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities
- Simon Willison, "Not all AI-assisted programming is vibe coding (but vibe coding rocks)," simonwillison.net, March 19, 2025, https://simonwillison.net/2025/Mar/19/vibe-coding/
- Andrej Karpathy, "Sequoia Ascent 2026 summary," Andrej Karpathy's blog, April 30, 2026, https://karpathy.bearblog.dev/sequoia-ascent-2026/
- "Andrej Karpathy," Wikipedia, accessed Nov. 2, 2026, https://en.wikipedia.org/wiki/Andrej_Karpathy
- "Hacker News," Wikipedia, accessed Nov. 2, 2026, https://en.wikipedia.org/wiki/Hacker_News
