Mention is not recommendation
1,658 videos have no recommendation label in the retained data. Reference, joke and critical mentions remain available when you choose All mentions.
An atlas of collective curiosity
An alternative view of HackerTube: the pages, the ontology beneath them and the knowledge graph that connects them, across 6,454 videos that keep finding their way back into the conversation.
Built on HackerTube, an experiment by Andrej Karpathy.
Select a cluster to explore its topics.
Inspect topic identity ↗KG curated by kg-generator, rdf-infographic-skill, and GPT-6 on behalf of Kingsley Idehen. Presentation redesigned by Claude Sonnet 5.5; data, links and graph unchanged.
What changes when discovery starts with people returning to an idea—and lets you inspect why?
The map shows where attention gathers. HackerTube itself shows the pages; this atlas also shows the ontology and knowledge graph beneath them, in three layers.
hackertube.org presents its collection as web pages: video cards, summaries and the Hacker News comments behind them. This atlas re-presents the same data through three connected layers: the HTML you read, the ontology that says what each thing is, and the RDF knowledge graph that makes every video, topic and recommendation addressable and queryable.
The “three eyes” are a way to read those layers: perception, intelligible structure, and integrated understanding. Each layer below is described in general terms, then as it appears in this collection.
“What am I looking at?”
HTML presents information through readable text, images, links, and interaction.
Most attention lands here.
This is the layer hackertube.org already offers, and the one this page also renders: cards, charts and a searchable catalog you can read and navigate.
“What does this mean, and what follows from it?”
An ontology defines concepts, relationships, and logical commitments, making interpretation and inference possible.
Less visible. Essential context.
Here, a video is a schema:VideoObject, a topic is a SKOS concept under one of 16 areas, and a comment is an annotation, so “recommendation” and “topic” carry defined meanings.
“How does this fit together, and how does it change my understanding?”
Exploring an RDF graph, guided by its ontologies and provenance, can help a person connect perspectives, examine assumptions, and recognize broader patterns.
High reuse value. Too often left in prose.
Every video, topic and retained mention is a linked, addressable statement, so “6,454 videos across 150 topics” can be queried and combined, not only read.
Take one claim this atlas can make: “Vectors | Chapter 1, Essence of linear algebra is a lasting recommendation.” Each layer lets you test it differently.
You see the video card and read a few of the comments.
You distinguish a recommendation from a reference or a joke, and a retained mention from any mention.
You follow the video to every recommending comment, the months they fall in and the topics it belongs to, then query the same evidence yourself.
Read, navigate and discover.
This HTML presents the story, charts, catalog and source evidence. Its prose communicates knowledge; the next two layers make meaning and reuse explicit.
Explore the HTML catalog ↓Name the entities. Define the relationships.
This collection combines Schema.org, SKOS, Web Annotation, PROV-O, RDF, RDFS, OWL and Dublin Core Terms, plus its own vocabulary. The topic taxonomy is instance data organized with SKOS.
Terms used by the companion graph. Select a term to inspect its meaning.
Query, connect and reuse.
In this collection, videos link to topics, topics link to broader areas, and comment annotations link evidence to videos. The Turtle companion carries these statements in reusable RDF form.
hackertube.org helps us perceive; the ontology helps us interpret; the graph lets us examine and query. This atlas keeps the three together, so every chart below can be traced back to its evidence.
Now the layers at work. The first question is what a mention actually says, because not every link is praise.
A link can recommend, cite, joke or criticize. Treating every mention as praise would erase the most useful distinction in the data.
40.6%of retained mentions carry a
recommendation label.
AI classifications of the retained export. Open the source comments to inspect the evidence.
Select a stance to highlight its videos throughout the atlas. Counts refer to comment-to-video mentions.
Once a mention has a label, time becomes the test: which recommendations come back, month after month?
One viral moment is easy to notice. A recommendation that returns across different months tells a different story.
Each dot is a video. Color identifies one of its topic areas. Small offsets separate overlapping values; the tooltip shows the exact record.
Patterns are only useful if you can act on them. The catalog turns the same evidence into a shortlist you can browse and filter.
The atlas and catalog are connected. Choose a topic, raise the evidence threshold, or search for an idea. Every card leads to the comments behind it.
Counts reflect the other active filters. Videos may belong to more than one topic.
The catalog is organised by sixteen subject areas. Open any of them to see how the topics beneath it are named and counted.
Choose an area to narrow the catalog. Category names link to their RDF identities.
How was this made, and where does it fall short? The method, its limits and the reusable recipe follow.
The source scans comments, retains recurrent video mentions, classifies intent, writes video cards, and ranks recurring recommendations. The RDF rendition preserves that pipeline and adds queryable identities and provenance.
1,658 videos have no recommendation label in the retained data. Reference, joke and critical mentions remain available when you choose All mentions.
The export retains at most 20 mentions per video, and comment text is cleaned and capped at 1,500 characters. The linked HN page is the original source.
Video summaries, stance labels, categories and why-HN explanations were generated by the source AI pipeline. The RDF records that provenance separately from the original comment identity.
The ranking rewards distinct recommendation months and gives extra weight to months with strong recommendations. Co-recommendation counts measure shared recommenders, not factual agreement.
Scan HN comments from 2007 through 2026 for YouTube links, retaining comments with at least 100 characters of their own text.
Select videos mentioned by at least three distinct people across at least three months; retain up to 20 mentions per video.
Read each comment with its HN story title and assign a recommendation, reference, joke or critical stance.
Combine retained mentions with video metadata to write summaries and explanations, then assign one to three taxonomy categories.
Compute recommendation months plus half the months containing strong recommendations.
This rendition converts the export to RDF and renders the catalog, comment evidence and graph views from that RDF. This final step is the rendition workflow, rather than part of the original extraction pipeline.
A video-discovery catalog assembled from YouTube links repeatedly discussed on Hacker News.
The 2026-10-31 export contains 6,454 video records and 39,834 labeled comment mentions, organized into 150 categories across 16 areas.
No. Only rec and strong_rec labels count as recommendations. Reference, joke and critical labels preserve other uses of a video link.
The source retained videos mentioned by at least three different people across at least three months, considering comments with at least 100 characters of their own text.
The score equals distinct recommendation months plus 0.5 times the number of months containing a strong recommendation. It rewards repeated recommendations over time.
No. At most 20 mentions per video are retained, favoring different commenters. Comment text is cleaned, link-normalized and truncated to 1,500 characters in the export.
The source credits Claude Sonnet 5.5 for comment classifications, video cards and topic assignments, coordinated by Claude Opus 5.5. These are AI-generated interpretations and can contain errors.
It records how many HN accounts recommended both videos. Each source list retains up to 20 related videos; missing links do not prove that two videos have no shared recommenders.
Stable entity identifiers and explicit relationships connect videos, categories, evidence, model-generated classifications and ranking measures. SPARQL can traverse and combine those relationships.
Yes. The exported video cards, taxonomy, comments and relationship counts support discovery and evidence browsing. A key is needed only for optional current YouTube metadata enrichment.
The export excludes YouTube API metadata such as durations, descriptions, view counts and publication dates. This rendition does not infer those missing fields.
The source releases the dataset under ODC-By 1.0 and requests credit to HackerTube and open-index/hacker-news. Comments remain the work of their authors.
No. Summaries and recommendation explanations describe the source video cards and HN discussion. The export does not include video transcripts.
This rendition needs no account. Its browsing filters remain in the page; opening a source video or comment follows the corresponding external service.
Source-exported comment text with HTML stripped, links normalized and a maximum length of 1,500 characters.
The source classification for criticism or debunking of a video.
Recommendation months plus half the months containing a strong recommendation.
A source-assigned classification indicating evergreen, dated or news content.
One of 150 export-specific topical groups arranged under 16 areas.
The exported title, channel, generated summary, why-HN explanation, classifications and ranking fields for one YouTube identifier.
A retained HN comment and its source URL together with a separately attributable AI stance classification.
The rec or strong_rec classification assigned to a video mention.
A calendar month with at least one retained rec or strong_rec mention for the video.
The source classification for a neutral citation of a video.
A source-exported comment-to-video record; no more than 20 are retained per video.
A topic-and-evidence lens: all 16 areas and 150 categories, leading recommended videos from each area, and selected supporting mentions. Core shows the most connected entities; Full expands this lens. The catalog and SPARQL recipes cover the complete dataset.
Inspect verified results bundled with this collection, or edit a recipe and open it on your SPARQL endpoint. Endpoint execution requires the companion Turtle to be loaded under the selected graph IRI.
PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
PREFIX ht: <https://x.com/karpathy/status/2116957135661945283#>
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
SELECT ?type (SAMPLE(?s) AS ?sampleEntity) (SAMPLE(?label) AS ?sampleLabel) (COUNT(?s) AS ?entityCount)
WHERE { GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/hackertube-knowledge-collection-gpt6-1.ttl> { ?s rdf:type ?type . OPTIONAL { ?s rdfs:label ?label } } }
GROUP BY ?type ORDER BY DESC(?entityCount)PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
PREFIX ht: <https://x.com/karpathy/status/2116957135661945283#>
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
SELECT ?videoIri ?title ?score ?firstYear ?lastYear
WHERE { GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/hackertube-knowledge-collection-gpt6-1.ttl> { ?videoIri rdf:type schema:VideoObject ; schema:name ?title ; ht:score ?score ; ht:strongRecommendations ?strong ; ht:firstMentionYear ?firstYear ; ht:lastMentionYear ?lastYear . FILTER(?strong > 0 && ?lastYear > ?firstYear) } }
ORDER BY DESC(?score) LIMIT 30PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
PREFIX ht: <https://x.com/karpathy/status/2116957135661945283#>
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
SELECT DISTINCT ?videoIri ?title ?recommendationIri ?criticalIri
WHERE { GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/hackertube-knowledge-collection-gpt6-1.ttl> { ?videoIri schema:name ?title . ?recommendationIri rdf:object ?videoIri ; ht:stance ?stance . VALUES ?stance { ht:stance-rec ht:stance-strong_rec } ?criticalIri rdf:object ?videoIri ; ht:stance ht:stance-critical . } } LIMIT 30PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
PREFIX ht: <https://x.com/karpathy/status/2116957135661945283#>
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
SELECT ?categoryIri ?category ?areaIri ?area
WHERE { GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/hackertube-knowledge-collection-gpt6-1.ttl> { ?categoryIri skos:broader ?areaIri ; schema:name ?category . ?areaIri schema:name ?area } } ORDER BY ?area ?categoryPREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
PREFIX ht: <https://x.com/karpathy/status/2116957135661945283#>
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
SELECT ?videoIri ?title ?relatedVideoIri ?relatedTitle ?shared
WHERE { GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/hackertube-knowledge-collection-gpt6-1.ttl> { ?linkIri rdf:subject ?videoIri ; rdf:predicate schema:isRelatedTo ; rdf:object ?relatedVideoIri ; ht:sharedRecommenders ?shared . ?videoIri schema:name ?title . ?relatedVideoIri schema:name ?relatedTitle . } } ORDER BY DESC(?shared) LIMIT 30