Argues that HackerTube's quality comes from aligning the ranking signal with the audience's actual taste — 'would a working engineer link this in a comment thread' — rather than optimizing watch time across a billion people. The point isn't that YouTube's models are bad, it's that their loss function is wrong for this specific audience, which is why a zero-ML recommender outperforms them here.
Notes that nothing in HackerTube is technically hard — the HN BigQuery dataset is public, YouTube URL extraction is a regex, and thumbnails come free from oEmbed. The value is in the editorial choice to wrap a curated corpus in a YouTube-shaped UI so the contrast between the two ranking substrates is immediately legible.
Karpathy shipped HackerTube as a weekend project with a one-sentence pitch — a grid of 6,454 videos pulled from 39,834 HN comments since 2007 — betting that the HN crowd would prefer their own collective link-sharing as a ranking signal over algorithmic recommendations. The 943-point HN thread validates the bet: the aesthetic landed precisely because it was the artifact the audience was quietly hoping someone would build.
Andrej Karpathy quietly shipped a weekend project called [HackerTube](https://hackertube.org/) and wrote it up on his bear blog. The pitch is one sentence long: it's a YouTube-style grid of 6,454 videos, every one of them pulled from a YouTube link that appeared in a Hacker News comment at some point since 2007. The source corpus is 39,834 HN comments containing a `youtube.com` or `youtu.be` URL.
There's no model in the loop. There's no personalization. There's no watch history, no engagement optimization, no "because you watched." The ranking signal is simply: did a human on Hacker News think this video was worth linking to in a thread with other humans who read Hacker News. That's it. The UI is deliberately YouTube-shaped — thumbnails, titles, a grid — so the contrast is visible at a glance: same shell, radically different substrate.
The HN thread hit 943 points fast, which is the kind of number you see when a project lands on exactly the aesthetic its audience was quietly hoping someone would build. There's nothing technically novel here. The BigQuery HN dataset is public, YouTube URL extraction is a regex, thumbnails come free from the oEmbed endpoint. The novelty is the taste.
The interesting thing about HackerTube isn't the artifact — it's what the artifact implicitly argues. Karpathy, who has spent the last decade building the models that power modern recommender systems, just published a counter-example: a recommender with zero ML in it that is probably better than YouTube's for a specific audience. Not because the model is bad, but because the loss function is wrong. YouTube optimizes for watch time across a billion people. HackerTube optimizes for "would a working engineer link this in a comment thread," and for the subset of videos that answer *that* question, the second signal is just higher quality.
This is a pattern worth naming. Over the last few years, several of the most-shared tools on HN itself have been variations on the same move: take an existing text corpus with implicit curation (HN comments, GitHub stars, arxiv-sanity bookmarks, Stack Overflow accepted answers), re-shape it into a consumption surface, and ship. [arxiv-sanity](http://arxiv-sanity-lite.com/), also Karpathy, did this for papers. [hntop](https://news.ycombinator.com/) is itself an instance of it. The common thread is that the curation already happened; the work is just surfacing it.
This matters right now because the industry's dominant narrative about discovery is going the other way. The pitch everywhere — Perplexity, ChatGPT with search, Google's AI Overviews, every "AI-native" news reader — is that a model should read everything on your behalf and distill it. HackerTube says the opposite out loud: a well-shaped reservoir of existing human judgment, exposed as-is, often beats a model trained to approximate that judgment. One is 400 lines of SQL and HTML. The other is a billion-dollar inference bill.
The community reaction in the HN thread leans into this. The top comments are mostly people naming the specific videos they're glad are in there — Jonathan Blow talks, old SIGGRAPH recordings, Rich Hickey's "Simple Made Easy," Bret Victor's "Inventing on Principle," the Casey Muratori handmade series. These are videos that YouTube's algorithm will not reliably surface to a developer in 2026, because they don't retain mass audiences, don't get re-upped, and lose to shorter content with better thumbnails. They stay alive because engineers keep linking to them in comments. HackerTube just made that invisible graph visible.
The practical takeaway isn't "go clone HackerTube." It's that if you're building any kind of discovery, search, or recommendation feature inside your product, there's a question worth asking before you reach for embeddings: what's the implicit curation signal already sitting in my database? Support ticket threads, PR comments, Slack message reactions, Linear issue links, the URLs your users paste at each other — these are curation events your team has been generating for free, usually for years, and almost nobody surfaces them.
A few concrete patterns the HackerTube approach makes obvious:
- Link-in-comment is a vote. If an engineer pastes a URL into a code review or a support thread, that's a stronger signal than a star, a bookmark, or a thumbs-up. It survived the friction of being typed into a conversation where other engineers would see it. Mine these before you mine anything else. - Shape matters more than ranking. HackerTube doesn't rank videos interestingly — it mostly sorts by recency or popularity on HN. The win is the *shape* (a grid of thumbnails) applied to the *corpus* (HN's video links). You don't need a better model; you need a better container. - Scope the audience, not the content. The reason HackerTube feels curated is that HN is a narrow audience with consistent taste. Any discovery feature you build will feel smarter if you pick a tight audience segment and over-index on signals from them, instead of trying to serve everyone with one ranker.
If you're on a product team and someone pitches "let's add AI-powered recommendations," the honest first question is whether you've exhausted the non-AI signals already in your warehouse. In most cases, you haven't. A SQL query against your own activity log, re-shaped into a browseable surface, will often beat a vector index on the same data — and it's debuggable, cheap, and doesn't hallucinate.
HackerTube will probably not be a product. It'll be a weekend project that stays a weekend project, which is part of the point — it works precisely because nobody has to monetize it, so nobody has to optimize it for engagement, so the signal stays clean. But expect to see more of this shape: small, static, scoped surfaces built on top of a decade of someone else's curation, shipped by people who understand that the hard part was never the ranking algorithm. The hard part was finding a corpus where the curation had already happened and then having the taste to leave it alone.
YouTube-style UI over 6,454 videos sourced from 39,834 HN comments since 2007.<p><a href="https://hackertube.org/" rel="nofollow">https://hackertube.org/</a>
→ read on Hacker NewsTop 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.