HackerTube in 2 Hours: Building an AI-Curated Content Layer Your Team Can Actually Copy

Most coverage of AI applications moves from a model release straight to a productivity claim. The HackerTube episode — put online recently at hackertube.org by Andrej Karpathy — is more useful read from the other end: as a worked example of swapping a ranking algorithm for a mined human signal, then letting an agent assemble everything else around it. Karpathy, once Tesla's AI director and one of OpenAI's co-founding members, says he needed about 2 hours with Anthropic's coding tool Claude Code to go from complaint to live website. The raw material was not viewing telemetry but the comment archive of Hacker News, and the result is a browsing site built on 6,454 videos that community members kept pointing at across the past 20 years, sifted out of 43 million comments. His summary on social media was blunt: the site offers a "clearly better watching experience" than the recommendation system it replaces.

Background: the cover image you distrust was optimized, not accidental

The irritation behind the project is widely shared and easy to dismiss as taste. It is more productive to read it as a system output. YouTube's own Studio tooling says as much: its Test and Compare feature, launched for thumbnails and expanded to titles in 2025, lets creators run up to three variants against a slice of their audience and keep the winner (Tubefilter, July 2025). An engagement-optimized packaging slot is the predictable result of that loop, which is why HackerTube's most quoted design decision — discarding creator-supplied cover art in favour of a frame captured near the middle of the video — is not a joke about clickbait but a control placed on a known distortion.

Karpathy's own write-up (karpathy.bearblog.dev/hacker-tube/) adds the funnel detail that the headlines usually skip. Inside that archive, roughly 40,000 comments actually carried a video recommendation; those condensed into a set spanning 2007 to 2026, grouped across 150 topics. Hold that against the 43 million comment figure and you get one surviving item per several thousand comments. Curation quality, in other words, is mostly the price of rejection — and rejection is the part most teams never budget for.

Inside the build: what the agent owned, and the four steps it did not

The setup is unremarkable and that is the point: Claude Opus 5.5, the 5.5 model line, with reasoning intensity set to Medium, running inside the Mac desktop app of Claude Code.

Work item	Performed by	Why it matters to your own pilot
Reading the comment corpus, deciding what counts as a recommendation worth keeping	Claude	This definition, not the code, sets the ceiling on final quality
Data pipeline, search, recommendation logic, page assembly	Claude Code	The mechanically tedious layer, fully delegated
Updating code, hardening security, deploying and tuning performance once a server address was supplied	Claude	He handed over the IP and asked for the rest
Deciding what to build and phrasing the next instruction	Karpathy	Conversation was his only working interface
Creating the API key for YouTube video metadata queries in the Google Cloud interface	Karpathy	Credentials stayed outside the agent
Spinning up the VPS that hosts the site	Karpathy	Infrastructure provisioning remained human
Buying the domain	Karpathy	Purchases and account ownership are not delegated

Reflecting on the split, Karpathy observed that many projects which once might have taken weeks or months can now be finished in a morning, after a handful of requests to an AI — and then asked the question that should be printed above every pilot review: "what other possibilities have we not yet thought of?"

A step-by-step template for your own recommendation-trace project

1. Locate a system that already records endorsement. Slack or Teams threads, wiki page comments, Jira issues that people keep linking to, answers in a support forum. What you want is evidence that someone vouched for an artifact, not evidence that someone clicked it. Slack's guidance on AI knowledge bases frames the same idea: the value sits in working with what teams already produce instead of commissioning fresh documentation (slack.com).
2. Define "sustained recommendation" before you prompt anything. Independent mentions, different people, spread across time — the target Karpathy described was continuity rather than a single shout-out. Write it as one rejection rule.
3. Cap the output. A shortlist of 6,454 items is a deliberate constraint. An unbounded "relevant results" list reproduces the original problem in a new colour scheme.
4. Strip the packaging. Replace promotional assets with neutral representations (a mid-video frame, a first-line abstract, a file's date of last human edit) so presentation cannot outvote substance.
5. Hand the boring half to a coding agent. Extraction, deduplication, search, page assembly, deployment scripts. Keep the definition and the acceptance bar in your own head.
6. Keep four things human. Credentials, data-access approvals, hosting, and domain or billing ownership — the exact list Karpathy executed himself.
7. Ship read-only and account-free first. HackerTube stores nothing but a local browser session state, which keeps privacy review small and makes the first release cheap to withdraw.
8. Publish the derived dataset next to the app. The downloadable ZIP of comment-analysis output turns one team's pass into everyone else's starting point; internal forks are how a curation layer compounds.

Common misconceptions to retire before you pitch this internally

• "One person did it fast, so our backlog is a weekend." The figure of approximately 2 hours belongs to a personal project with no users to onboard and no compliance perimeter. It does not mean similar products move at the same speed through data processing, long-term maintenance or security assurance.
• "The agent deployed it, so the agent secured it." Anthropic's own Claude Code security documentation asks teams to work with least privilege, reviewable approvals and explicit guardrails (code.claude.com). A deployment finished in minutes is an exposure surface opened in minutes if a live key was pasted into the conversation.
• "Curated by humans means unbiased." The pool inherits the archive's blind spots: Hacker News is an English-language, technology-heavy community, and 6,454 videos from the past 20 years cannot stand in for a whole content universe. Know whose taste you are automating.
• "You need a recommender model." HackerTube ships no trained ranking model at all. For most internal use cases, a proxy signal plus a rejection rule is cheaper, auditable and faster to explain to skeptical stakeholders.
• "This competes with the hosting platform." The videos stay where they are; the build only queries YouTube video metadata via a Google Cloud key. What was rebuilt is selection, not infrastructure — which is precisely why one person could afford it.

Trend read: curation becomes buildable, and the bottleneck moves to judgement

Two signals point the same direction. First, where agent usage actually concentrates: Anthropic's Economic Index reported that 44% of API traffic in its enterprise sample mapped to computer and mathematical tasks, against 36% of Claude.ai usage (anthropic.com/research). Implementation work is the first place delegated effort lands, and HackerTube is that pattern shrunk to individual scale. Second, what becomes scarce when implementation is cheap: taste, access rights to the underlying archive, and the willingness to discard almost everything. When a build costs a morning, no part of the advantage came from building it.

The more interesting consequence for knowledge work is structural. Large platforms optimize distribution toward the average viewer because that is what scale monetizes. A personal curation layer does the opposite: it is allowed to be narrow, and narrow is what expertise looks like. Expect this category to fill with vertical forks — 150 topic groupings is already a shelf of candidate niche products — and expect incumbents to respond on the packaging side rather than the signal side, since their revenue is attached to attention.

What to do this week

• Pick one artifact stream your team already distrusts: the "most popular" internal wiki list, the sales enablement folder, the training catalogue.
• Ask the data owner for endorsement traces (who linked this, who cited it twice, who sent it to a newcomer) instead of click logs.
• Write the rejection rule in a single sentence and have a domain expert sign it before a prompt is typed.
• Treat the pilot as throwaway, and budget the security review, the credential boundary and a named maintenance owner separately.
• Choose the success metric up front: how many of the top 50 items an expert recognizes as worth their time, not how many were served.

Read as an application pattern, HackerTube's transferable core is the inversion: mine what people have already vouched for as the dataset, and let the agent build the interface on top of it. The roughly 2 hours Karpathy reported should stay labelled as individual practice, not a promise about data handling, upkeep or protection at organizational scale. His closing question — "what other possibilities have we not yet thought of?" — is best answered here with something deliberately unglamorous: the curated shortlist your company never built because the effort-to-credibility ratio was too high. That ratio is what just changed, and the teams that notice will simply stop settling for the default ranking of their own knowledge.