Research · 2006–2026 · 49,324,111 items
The anatomy of Hacker News: 18 findings from 49.3M items
Not what is on HN, but how HN works: content half-lives, the collapse of discovery lag, attention concentration, the physiology of discussion, language turnover — and what a community pays for stability with churn.
Every chart is computed from DuckDB by deep_analysis.py and is reproducible; data caveats at the end.
words and deleted columns (coverage 0% / 2.3%). Frequency charts for those years fall back to text regexes; deletion rates are marked unobservable. words is an alphabetically sorted token set, not a sequence. 2026 covers January–August only. These choices apply throughout; the charts do not repeat them. On narrow screens, charts scroll horizontally to keep axis labels legible.Content half-lives differ by 8×
For every object mentioned by at least five distinct people, we measure the span between its first and last appearance on HN. The median wiki entry spans 3,085 days (8.4 years) and 38.6% of them survive a decade; the median blog post spans 620 days, and only 1.5% live ten years. One feed contains both a codex and a flyer.
Median span in days (bars); the right-hand note is the share of objects with a 10+ year span. Metric: site_obj, mentions ≥ 5.
Chasing new, digging old
How old was an object when a link to it was posted, measured from its first HN mention? In 2007, 93% of links pointed at objects first seen within 30 days. In 2025 that share is 47% — while links to 10+ year-old objects rose from 0% to 15.3%. The two directions are not contradictory: the radar got faster and the archive got thicker.
All families, by link year. 2026 is a partial year.
Split by family, the two time economies are nearly incompatible. In 2024–26, 92.7% of blog links pointed at something first seen within a year; for wiki it was 26.6%, while 58.3% of wiki links referenced objects at least five years old.
| fam | ≤12 months | 5+ years |
|---|---|---|
| music | 94.9% | 1.4% |
| blog | 92.7% | 3.1% |
| build | 92.6% | 6.9% |
| incident | 90.5% | 4.8% |
| dataset | 89.9% | 0.7% |
| story | 83.9% | 16.1% |
| patent | 81.8% | 6.7% |
| paper | 81.4% | 5.4% |
| career | 77.5% | 22.5% |
| video | 76.0% | 11.2% |
| podcast | 69.0% | 6.0% |
| film | 63.7% | 24.4% |
| repo | 62.4% | 17.0% |
| book | 57.8% | 29.7% |
| course | 56.1% | 27.6% |
| standard | 40.6% | 33.7% |
| wiki | 26.6% | 58.3% |
Sleeping-beauty papers: from 9 years to zero
Using the first four digits of arXiv identifiers to recover publication dates (26,493 papers), we measure the gap between publication and the first HN link. A 2007 paper waited a median of 105 months, and 68.9% waited six years or more. For 2025 papers, 97.6% appear within a year. HN stopped being an archaeology crew and became a news ticker.
Grouped by publication year. The ≥36-month sleeper share (dashed) collapses in parallel. Recent low values include right-censoring.
Canonical sleepers include Legg & Hutter's definition of intelligence (published 2007, shared 8.8 years later), an introduction to quantum computing (8.0 years), and “Rethinking Set Theory” (6.5 years). Their common trait: they were rediscovered by a thread, not pushed by news.
caveat Papers published after 2023 have not had time to demonstrate a five-year rediscovery, so recent low values are partly right-censored. The 2007–2019 decline is fully observed and the direction is reliable.
Attention Gini: tools became more concentrated
Within each family, how large a share of recommendations (distinct recommenders) goes to the top 1% of objects? Across 2021–26, the top 1% of repositories (1,491 repos) captured 17.7% of repo recommendations, up from 13.4% in 2007–15. Patents went the other way: 36.4% → 7.1% as the object count grew from 304 to 1,080.
Three eras per family: 2007–15 / 2016–20 / 2021–26. Values are the recommender share of the top 1% of objects.
Co-citation: how content families meet
Two different families appearing in the same story or comment. video × wiki is the largest pairing (8,129 times): a video explains, an encyclopedia entry verifies. Only 1.3% of object-bearing items contain more than one family — HN's multimedia mixing is rarer than it feels.
Top 10 co-occurrences (family A × B in the same source item).
The paper × repo pairing deserves its own line. Items containing both an arXiv paper and a GitHub repository grew from 53 in 2016 to 244 in 2023 (4.6×). On HN, research is now discussed as a paper-plus-reproduction bundle.
paper × repo co-cited items, by first appearance. 2026 is eight months.
Weekend HN is smaller, but worth more
Weekends produce only 19.1% of submissions (28.6% of the calendar is weekend), yet weekend stories average more points: 12.9 vs 11.2, median 2 vs 1. The emptier front page belongs to encyclopedias, film and video; courses and standards retreat.
Weekend share is 19.1%; mean points 12.9 vs 11.2.
From a 24-hour sprint to a week-long marathon
In 2010, 97.7% of comments arrived within 24 hours of the parent story. By 2025 it was 84.6%. The 2–7 day window grew from 18% to 30%, and comments arriving a week or later quadrupled. A front-page discussion now lives about a week instead of a day.
Comment lag relative to the parent story (all 43M comments). 2026 is eight months.
caveat The 24h and 2–7 day columns overlap in definition. Read the trend, not the sums. The mechanisms (larger threads, second-chance re-ranking, timezone spread) cannot be separated here.
Shorter comments, fewer citations
2021 → 2026: median comment length fell from 35 to 30 words and the mean from 45.3 to 39.9 (−12%). The share of comments containing an external link fell from its 2015 peak of 14.5% to 9.2% (−37% relative). This is not “the comments got worse”; the genre shifted from citation culture to conversation culture.
Share of comments containing at least one link (links.kind=comment / all comments).
caveat The words column is empty for 2024–25, so the length series only covers 2007–2023 and 2026. Link density is unaffected.
From bungalows to deep wells
Top-level comments (direct replies to a story) fell from 47.9% of all comments in 2007 to 20.3% in 2021. Replies six levels deep grew from 4.4% to 14%. At constant volume, HN turned from “many people saying one thing” into “fewer people answering each other”.
Share of comments by reply depth; depth 6+ is merged. 2026 is partial.
2025 was Rust's year
Mentions per 10,000 comments. Rust climbed from 0.4 (2007) to 89.4 (2025), almost catching Python (92.6); in the first eight months of 2026 it overtook it. Zig rose from 0.2 to 15.6. JavaScript fell from 106 (2011 peak) to 42.8, and Ruby/PHP now sit at 15–17% of their 2007 levels.
Comments mentioning each term per 10,000 comments. 2024–25 use the text fallback; 2026 is eight months.
caveat “go” is an English verb and “c” is a letter; neither is a language signal. The golang column counts only the explicit spelling, so Go's real mindshare is underestimated. That is a limit of token counting, not a data error.
Crypto's 4-year half-life, AI's 14× climb
Narrative words have clocks. bitcoin fell from 80.4 (2021) to 9.1 (2026); blockchain from 30.5 to 5.0; web3 from 11.6 to 1.0. Crypto vocabulary has a half-life of roughly four to five years. AI went the other way: from 50.6 to 729.3 (14.4×), and 7.3% of 2026 comments mention AI.
AI vocabulary (log scale, axis floor 0.1). Series begin at their first mention; zeros are omitted rather than drawn on the floor. claude went from 2.8 to 204.4 (~73×) and overtook every other model brand in the first eight months of 2026.
Crypto narratives (linear). “crypto” stays high because it also means cryptography — the semantic drift is itself a signal.
Three moves of personal publishing
Share of links by domain. 2007–12 was the Blogspot era; Medium peaked at 2.10% in 2018; Substack took over from 2019–22 (0.61% in 2022, first beating Medium's 0.57%); by 2026 Substack reached 1.06% and github.io 1.33% — author-owned assets together now exceed Medium's peak. The Twitter-to-X domain swap is equally visible: x.com is already 1.8× twitter.com.
Percentage of all links that year. twitter/x and blogspot/wordpress are merged series.
Show HN doubled, Ask HN halved, AI in one title in ten
After 2023, Show HN rose from 3.5% to 9.7% of stories, while Ask HN fell from 5.1% in 2010 to 1.5%. Titles containing “AI” reached 10.3% of all stories. HN moved from “reading news / asking questions” to “shipping work + long comment threads”, and AI stopped being a topic and became the background.
Share of story titles matching each pattern. 2026 is partial.
From eight jobs per seeker to 0.6
The three monthly threads are a ready-made business-cycle index. The jobs/seeker comment ratio peaked at 8.0 in 2018, flipped in 2023 (0.8), and fell to 0.6 in the first eight months of 2026 — supply is now 1.7× demand. Meanwhile the average hiring post shrank from 298 to 103 comments: fewer roles, and thinner ones.
Hiring-thread comments / job-seeker-thread comments. Dashed = parity (1.0). Series starts in 2014, when the seeker thread first appears.
Halved retention, constant 1% power
Among authors active in their first year, the share still commenting a year later fell from 43.6% (2007 cohort) to 24.9% (2018); five-year survival fell from 31.2% to 13.0%. Yet the top 1% of authors still write 35–40% of all comments (39.7% in 2025), and about four in ten active authors each year are new. HN is “1% resident + 40% flow”: a funnel, not a club.
Cohort survival: share still active 1 / 5 years after the first comment year.
Top-1% authors' share of comments, and newcomers' share of active authors.
2026's moderation shadow is the heaviest since 2011
In the first eight months of 2026, 26.5% of stories were dead (flagged or sunk) — the highest since 2011 (31.25%) — and 4.02% of comments were dead, an all-time high for this sample. User-initiated deletion stayed stable at 2.5–3.3%, so the increase comes from community/system suppression, not regret. One story in four now sits below the waterline.
Dead share by year. The deleted column is missing for 2024–25, so this chart uses dead only; 2026 is eight months.
New species, and 3,600 fossils
Objects first seen since September 2024, ranked by distinct recommenders over the last 18 months. Twenty of the top 25 are repositories, and 12 are directly about AI coding or agents. Non-AI winners cluster in developer experience (jj, ghostty, apple/container) and anti-AI infrastructure (Anubis' “Proof of React”, Wikipedia:Signs_of_AI_writing).
| # | object | recs / 18mo | first seen |
|---|---|---|---|
| 1 | anthropics/claude-code repo | 360 | 2025-02 |
| 2 | ggml-org/llama.cpp repo | 142 | 2025-02 |
| 3 | openai/codex repo | 139 | 2025-04 |
| 4 | google-gemini/gemini-cli repo | 93 | 2025-06 |
| 5 | ghostty-org/ghostty repo | 79 | 2024-12 |
| 6 | anomalyco/opencode repo | 75 | 2026-01 |
| 7 | apple/container repo | 74 | 2025-06 |
| 8 | Wikipedia:Signs_of_AI_writing wiki | 73 | 2025-07 |
| 9 | TecharoHQ/anubis repo | 60 | 2025-03 |
| 10 | obra/superpowers repo | 55 | 2025-10 |
| 11 | jj-vcs/jj repo | 52 | 2024-12 |
| 12 | badlogic/pi-mono repo | 50 | 2025-11 |
On the other side: 3,600+ objects first seen by 2016, mentioned at least 15 times, and silent since 2023 — 1,561 repositories and 1,461 wiki entries among them. Each fossil marks a slice of technical history and is ready-made for an archive page.
Silent objects (seen before 2016, mentions ≥15, no mentions after 2023).
From generalists to monoculture?
How many families does one author link to in a year? Breadth peaked in 2022 (1.62 families on average; 35.3% crossed two or more), then fell for three straight years to 1.48 and 28.7% in 2025. This may be the first quantitative signal of AI-era attention monoculture.
Authors who posted at least one object link that year. 2026 is partial and omitted.
Host structure: the face of AI infrastructure
In 2024–26, the dataset family is already a HuggingFace monopoly (2,040 links). arXiv remains the largest paper entrance (24,533), and MIT OCW leads courses (262) ahead of Coursera (149). Build a vertical catalogue by getting these entrances' metadata right first.
| fam | top host | links |
|---|---|---|
| book | www.goodreads.com | 2,081 |
| book | www.amazon.com | 2,033 |
| book | books.google.com | 597 |
| book | openlibrary.org | 173 |
| course | ocw.mit.edu | 262 |
| course | www.coursera.org | 149 |
| course | www.udemy.com | 128 |
| course | www.khanacademy.org | 99 |
| dataset | huggingface.co | 2,040 |
| dataset | www.kaggle.com | 180 |
| dataset | paperswithcode.com | 135 |
| dataset | data.gov | 4 |
Ten judgments
- Band content by shelf life. Half-lives differ 8×; one “last N years” filter destroys evergreen categories.
- HN is getting both faster and more archival. 47% of links point to objects ≤30 days old; 15% to objects older than ten years. Both labels are half-true.
- Paper discovery lag is gone. For a paper site, freshness is the product; “re-reading classics” is a different product.
- AI is context, not topic. 7.3% of comments and 10.3% of titles mention AI; the signal now lives one layer down (agent / model / tool).
- Rust went mainstream. 2025 density 89.4 vs Python 92.6, ahead by the first eight months of 2026; Zig (15.6) is the second signal of the same infrastructure migration.
- Narratives halve in 4–5 years. Crypto vocabulary fell to one-tenth within 4–5 years of peak; historical baselines beat gut feel for “the next hot word”.
- Discussion shape changed. Comments −12% shorter, links −37%, depth 6+ from 4.4% to 14%. Citation culture gave way to conversation culture; object extraction from comment text gets thinner every year.
- The labour market flipped. Jobs/seeker ratio 8.0 (2018) → 0.6 (2026): one of the cleanest public indices of the tech job cycle.
- Stability comes from churn. Retention halved, yet ~140k monthly authors stay constant, 40% of them new; the top 1% always writes a third or more.
- Treat the dead zone as a separate sample. 26.5% of 2026 stories are suppressed; ignoring the sunk quarter overstates consensus.