Andrej Karpathy Built His Own YouTube in About Two Hours

He built his own YouTube in about two hours

Andrej Karpathy went to YouTube late on Friday night and hated what he saw. He called the home page a wall of slop in a post on his own blog. The next morning, he asked Claude Opus 5.5 to build him something better.

The result is HackerTube, a video site built only from videos people recommended on Hacker News. Hacker News is a long running discussion board for tech readers. Opus pulled the board's comments, filtered them, and used Sonnet to cut the spam.

The pile narrowed to about 40,000 comments that recommend a video. From those came 6,454 videos across 150 topics. The recommendations span almost twenty years of readers pointing at videos they loved.

The site skips the uploaded thumbnail and shows a frame from the middle of the video instead. Recommendations work without an account, and your place is kept in your browser. You can also export the data and build your own version.

Karpathy says his own share of the work was small. He made an API key, started a server, and bought the domain name. Claude built, deployed, and secured the site, and the whole build took about two hours.

The point here is not that YouTube should worry. One person can now build a personal version of a big service in a morning. The hard part was never the code in this story. It was years of people saying, watch this one.

One shared line for saying no to AI training

Every website carries a small text file called robots.txt. It holds rules for automated visitors, like search crawlers and AI bots. Until now, a site that did not want its pages used to train AI had to write a separate rule for each company.

On October 30, Google, Microsoft, and OpenAI said they will honor one shared group in that file. Search Engine Journal reported the announcements. The group uses a shared name, __usage_ai-training, and then the normal Allow and Disallow lines a site already knows.

The idea came from Kevin Kelley at Anthropic. He sent it to the AI Preferences working group at the IETF, which handles internet standards work. Anthropic has started building support for it. Meta expressed support too, and gave no date.

The start dates differ, and none of the three honors the line yet. OpenAI aims for the coming months, and Google aims for early 2027. Microsoft starts on February 1, 2027, and adds a page level choice for grounding. Grounding is when an AI answer pulls facts from a live page.

Two limits matter here. The line covers future training only, and it does not stop a company from visiting a site or listing it in search. It is also voluntary, and the proposal is still a draft, not a finished standard.

Try this if you run a website. Open your robots.txt file this week and read what it says today. You can add the shared group now with a Disallow line, and keep your current company rules in place too. The shared line will start to count as each company turns its support on.

Agents have placed only a few hundred DoorDash orders

We hear a lot about AI agents shopping for us. DoorDash put a real count on it this week. On its earnings call, the company said third party agents have placed only several hundred orders in total. TechCrunch reported the comments.

Set that count next to a normal DoorDash day. The company said in its 2025 report that it finishes about 3 million deliveries a day. Several hundred agent orders is a rounding error beside that.

Chief executive Tony Xu kept both doors open in his answer. He said you never want to fight gravity, and you also do not want to be delusional. Most people still order in the DoorDash app, and the company wants to learn how agent ordering should work.

DoorDash is not sitting this out either. It has its own text to order agent, a chatbot called Ask DoorDash, and tools that let other chatbots place corporate orders. Other stores are split on the idea. Amazon blocks personal agents, and Walmart welcomes them.

So agents can place an order, and the demo works. Shoppers just have not moved yet. That gap between can and do is where most agent stories live right now.

Also today, OpenAI is changing how paid ChatGPT use is counted. Starting in December, Plus, Pro, and Business plans will share one allowance across ChatGPT and Codex, OpenAI says in a help page. Quick Instant chats will not count, and OpenAI has not shared the size of the allowance yet. Apple has hardware landing this week too, as its M6 MacBook Pro and iMac arrive on November 6. The All-In Podcast spent its latest episode on Anthropic's filing to go public and what the numbers say about demand for AI.

Today, one reader built his own video site, websites got one shared way to set a boundary, and shoppers kept ordering the old way. The tools keep getting stronger. Our choices about how to use them are getting clearer too.

About this brief: This brief was researched and drafted by my personal AI assistant and approved by me before publishing. Product details change fast; check primary sources before you build on them.

Want more with the Tech Simplified crew? Subscribe on Substack.

Get the week's briefs in one email. The Briefcase packs Monday to Friday's AI Briefs into one Sunday email.

Follow me on LinkedIn for project updates, ideas for using AI, and lessons you can bring to your own work.