Andrej Karpathy says he had Anthropic’s Opus 5.5 build HackerTube, a YouTube-style site made only of videos that Hacker News users recommended, in about two hours, according to his blog post dated 31 October. It holds 6,454 videos drawn from about 40,000 recommending comments posted between 2007 and 2026.
That is about 6.2 recommending comments per video, by our arithmetic on his figures. Karpathy lists three things he did by hand: creating a Google Cloud API key, booting up a server and buying the domain.
Short explainer · Karpathy’s HackerTube
A vertical news Short summarising this story. Figures come from Andrej Karpathy’s HackerTube blog post; our arithmetic is labelled.
Business Pill · THE MODEL AND THE HARNESS
A short explainer on models and harnesses. It teaches the general idea only and says nothing about HackerTube.
The key insight: As we read Karpathy’s post, the notable detail is the default, not the model: his stated default is Opus 5.5 on Medium effort in the Claude Code desktop app, and lists three manual steps for a site he says took about two hours.
Karpathy writes that he went to YouTube on a Friday night and found “an enshittified wall of slop”. The next morning he gave Opus 5.5 the task: “Let’s write our own YouTube clone based on whatever videos have been recommended on Hacker News”.
By his account, Opus downloaded Hacker News comments, filtered them and “ran Sonnet over them in a Workflow (Claude Code feature)” to remove spam, narrowing the set to 40,000 comments that recommend a YouTube video. The result spans 150 topics.
He lists the site’s features. It uses the 50% thumbnail instead of the uploaded one and offers recommendations without user accounts, keeping state in the browser session. It has keyboard shortcuts, and it exports its data so readers can give it to their own agent to build their own version. “This takes ~2 hours now”, he writes.
Andrej Karpathy’s HackerTube, by his account: about 40,000 Hacker News comments that recommend a YouTube video, 6,454 videos from 2007-2026 across 150 topics, built with Opus 5.5 in the Claude Code desktop app in about two hours (Karpathy blog post, 31 October 2026). Comments per video: our arithmetic.
What Was Posted on X
Andrej Karpathyโs setup from his HackerTube blog today:
– โI use Claude Code (desktop app for Mac).โ
As Iโve been saying, terminal is the wrong interface for coding agents. Desktop apps are the best interface today.
– โI use Opus 5.5 on Medium effort as my default.โ
Thereโs no point to use the smartest model at the highest reasoning effort for most tasks.
Investor and founder Yuchen Jin’s 1 November post quotes two lines from Andrej Karpathy’s HackerTube blog post: “I use Claude Code (desktop app for Mac)” and “I use Opus 5.5 on Medium effort as my default.” The comments about terminals and reasoning effort in the post are Jin’s own views.
His Setup, in His Words
In an edit to the post, Karpathy describes his process: “I use Claude Code (desktop app for Mac). I use Opus 5.5 on Medium effort as my default.” He adds: “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.”
Once the server was up, he writes, “I gave Claude the IP and told it update, deploy, secure, harden and optimize the code, and it did the whole thing.”
Investor and founder Yuchen Jin, who highlighted the setup on X, added his own view that the terminal is the wrong interface for coding agents and that most tasks do not need the top model at the highest reasoning effort. Those are Jin’s opinions, not Karpathy’s.
As we read it, the detail that matters for buyers is the effort setting. Karpathy, a former OpenAI founding member and Tesla AI lead, says his default is Opus 5.5 on Medium effort, and describes a two-hour build where his own work was three account and hosting steps. For subscription and API plans priced on usage, the effort tier is the lever a heavy user controls.
The Structural Read
As we read it, the build delegated the filtering step too: by Karpathy’s account, Opus ran Sonnet over the comments in a Claude Code Workflow to remove spam.
As we read the numbers, about 40,000 recommending comments became 6,454 videos, about 6.2 comments per video by our arithmetic, across 150 topics and 2007 to 2026.
As we read the edit, Karpathy’s own role was account and hosting work: an API key, a server and a domain, with Claude doing the deploy, security hardening and optimisation he asked for.
Andrej Karpathy, HackerTube blog post, 31 October 2026
“I only talk to Claude and have it do everything.”
Three Implications
FOR CODING-AGENT BUYERS Karpathy says his default is Opus 5.5 on Medium effort in the Claude Code desktop app.
FOR THE WORK ITSELF He lists three manual steps for HackerTube: an API key, a server and a domain.
FOR READING THE CLAIM The two hours, 40,000 comments and 6,454 videos are Karpathy’s own figures; his post gives no costs or token counts.
The analysis puts it in one line: “The surface, not the model, governs autonomy.”
As we read it, Karpathy names the surface first, the Claude Code desktop app, and then the model and effort setting. The framework was published in July 2026 and does not discuss HackerTube.
The two-hour figure, the 40,000 comments and the 6,454 videos are Karpathy’s own statements; we did not audit the site’s data. His post does not give token counts, costs or which plan he used. Yuchen Jin’s comments are his own views.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Andrej Karpathy says Opus 5.5 built and deployed HackerTube, a site of 6,454 Hacker News-recommended videos, in about two hours, with three manual steps on his side. His stated default setup is Opus 5.5 on Medium effort in the Claude Code desktop app. That is about 6.2 recommending comments per video, by our arithmetic.
A note on sourcing. Every figure about HackerTube comes from Andrej Karpathy’s own blog post of 31 October 2026; we did not audit the site’s data. Yuchen Jin’s comments are his own views. The comments-per-video figure is our arithmetic on Karpathy’s numbers.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
