Daily · Article 01 · 5 min read
- Anthropic
Karpathy explains HackerTube: the AI work and the resources people prepared
Follow-up to original dated October 31; added explanation checked November 2; edit time unspecified
Summary
This November 2 follow-up checks Karpathy’s October 31 HackerTube post and added process account. It explains selection criteria, human-prepared access/server/domain resources and the AI’s development work.
When a video feed stops helping you find a worthwhile talk, one option is to start from recommendations made by people you trust. AI researcher Andrej Karpathy's HackerTube turns that choice into a small web service.
Key facts
In an October 31 post, Karpathy described building HackerTube with Claude Code, an AI tool that performs development work. This follow-up checks the original and its added process explanation on November 2. The added paragraphs do not show a separate edit timestamp. Author's post
His account reports roughly 40,000 recommending comments, 6,454 videos, 150 topics and about two hours of work. Those are figures from his build, not a fresh count of today's service or a reproducible time estimate for every reader. He does not give a total operating bill.
1. Choosing where recommendations begin
Hacker News is a community discussing technology. Karpathy says he extracted YouTube recommendations from its comments and used another AI model, Sonnet, in classification. The work organizes existing videos and a browsing interface; it does not create new videos.
A comment contains a recommendation; a video is the object it points to. Several comments can point to one video, so dividing comments by videos cannot establish the share of good videos. This diagram groups the stages in the author's account.
%% From recommendation material to a catalog · Karpathy's process account
flowchart TD
A["Hacker News comments"] --> B["Filter and classify video candidates"]
B --> C["Organize existing videos and topics"]
C --> D["A page for browsing and choosing"]The approach lets an operator choose a recommendation source. One community's interests can also skew the selection; it is not a quality evaluation covering every viewer.
The cat asks Q&A
The researcherWhat should I tell an AI before asking for a similar catalog?
Pocket NotebookDefine the audience and the material eligible for inclusion. In a fictional library-talk catalog, you might start with trusted organizers and the intended age group. Asking only for a page design leaves you without a way to judge why a video was included.
2. From human-prepared resources to AI server work
In the added explanation, Karpathy says he created a Google Cloud API key, started a VPS and bought a domain. An API lets one program request information from another service; an API key supplies access information for those requests. YouTube video metadata includes descriptions such as titles and duration, rather than the video file itself. YouTube video resource
A VPS is a rented computer kept available online; a domain is a name through which people find the service. The author says he prepared resources between the AI's work, then passed the running server's destination to Claude for updating, deployment, hardening and optimization. The figure separates overlapping preparation from the handoff to server work.
Resource preparation and AI development overlap
Person · prepare resources
- Create an API key for video information
- Start a rented server (VPS)
- Buy a domain
AI · data and interface
- Collect comments; extract and classify recommendations
- Build the video catalog and web interface
Server ready: the person shares its destination with AI
AI · work on the prepared server
Update, deploy, harden and optimize
Output · HackerTube
A web service for finding and choosing existing YouTube videos
The cat asks Q&A
- The researcher
Did the person also handle the code and server commands directly?
Pocket NotebookThe author says he delegated by talking to Claude instead of reading or writing code or running terminal commands himself. His manual work was preparing the API key, server and domain between the AI's work. Applying that division to your own task still requires decisions about accounts, permitted changes and spending limits.
3. Where state lives without a user account
Karpathy describes storing state in the browser without separate user accounts. That describes sign-in, not a guarantee of no communication with servers or video providers. Ordinary web pages request material from servers. How HTTP requests work
Browser-held state and the server's video catalog can have different owners and lifetimes. For example, a browser's localStorage is scoped to a site and can be removed when its site data is cleared. That is a general explanation of browser storage, not an audit of all HackerTube storage or tracking. Browser storage reference
| Item to examine | Reader's question |
|---|---|
| Personal likes or watch-later choices | Do they move to another device, and can they be exported or restored? |
| The service's video catalog | What are its sources and update date? |
| Actual playback | Which provider's player and policies apply? |
- Item to examine
Personal likes or watch-later choices
- Reader's question
- Do they move to another device, and can they be exported or restored?
- Item to examine
The service's video catalog
- Reader's question
- What are its sources and update date?
- Item to examine
Actual playback
- Reader's question
- Which provider's player and policies apply?
The cat asks Q&A
- The researcher
What should I preserve before changing devices?
Pocket NotebookFirst check whether exporting your choices and exporting the service's catalog are the same feature. If personal state can be exported, obtain the restoration steps too. No account does not by itself promise synchronization or backup.
4. Operating records after the first build
Once a first screen works, data updates, broken videos, costs and incident response remain. In the fictional library catalog, someone still corrects categories, handles deleted links and owns next week's update.
| Check | Example evidence |
|---|---|
| Was the intended material selected? | Inclusion rules and a reviewed sample |
| Is it ready to show publicly? | Titles, links, categories and rights conditions |
| Can it keep running? | Update owner, cost records and recovery/contact procedures |
- Check
Was the intended material selected?
- Example evidence
- Inclusion rules and a reviewed sample
- Check
Is it ready to show publicly?
- Example evidence
- Titles, links, categories and rights conditions
- Check
Can it keep running?
- Example evidence
- Update owner, cost records and recovery/contact procedures
The cat asks Q&A
- The researcher
Where should I start after receiving a finished message?
Pocket NotebookOpen representative lists and links at the actual public address, then obtain the update procedure. Read first-build speed alongside the continuing work to understand the delegated scope. The useful lesson is how human choices about material and permission connect to an AI's work.
To close
A small service shaped by your selection criteria
HackerTube illustrates a choice about recommendation material and navigation around existing videos. Before ordering a screen, an operator can decide which sources to trust, what execution to permit and who will keep checking the result.
Sources
- Author's postkarpathy.bearblog.dev
- YouTube video resourcedevelopers.google.com
- How HTTP requests workdeveloper.mozilla.org
- Browser storage referencedeveloper.mozilla.org