The board: twelve terms, three years
What everyone talked about, month by month
Each circle is a term, sized by its share of that month’s talk. Pick whose talk to show, then slide through three years or press play.
What pops next, and what holds
Why: spikes fade, slow builds stay
How we know: every term that peaked, and how far it fell
Do statements predict the crowd?
The stories behind the calls
What the labs talk about that the public doesn’t
Any term. Any voice.
Pick a term, then choose whose talk to follow. Each line is one source over time, drawn against its own peak.
The full record
The analyses and tools behind the findings above, folded away. Open any one.
Rising and fadingEvery term’s last 13 weeks against the 13 before.
What moves termsThe biggest weekly jumps and the launches behind them.
The biggest jumps, and what was happening
Events that moved several terms at once
Every launch week, and the steepest falls
Launch weeks
Falls
Lab vocabulariesThe words each lab uses far more than the others.
Which lab named a term first, and each lab’s share of headlines
Who goes firstWhen lab staff, outside voices and Hacker News took off on each term.
Six terms the record caught early, with the first statements
Talk vs useTalk peaks, installs keep growing.
Every term, every monthEvery term as a heat map.
Each row is a term and each cell a month, colored from the term’s quietest months (dark) to its own peak (bright). Switch the measure to compare Hacker News with lab staff, outside voices, Bluesky or Wikipedia; the rows stay in place.
Hover or tap a cell for the numbers. Tap a word to open its full record further down; the highlighted row is the word open there.
Look up a termEvery signal and statement for one term.
The full record for one term: every signal on one chart, what each group said about it, the lab headlines, and the moments that pushed it.
Every signal, one chart
Every line is drawn against its own peak (100), so you compare when things rose and fell, not their size. Statement lines count what lab people and outside voices said, smoothed over about three months; switch to dots to see each month’s count. Ticks along the bottom mark lab headlines; numbered pins mark key moments.
Show the numbers
What each group said
Key moments
People and momentsEveryone in the record, and every key moment.
Everyone in the record and every key moment, each linked to its public source.
The launches, essays, interviews and podcasts that pushed a word into circulation, each linked to its source. Tap a word to open it.
How it was measured
How to read this page
- Hacker News is the crowd. For every word, the share of active Hacker News users who used it each week and month. It is the longest, most even record of developer talk, so it sets each word’s status and peak.
- Lab people and outside voices are a curated record, not a count: dated, linked statements by people at OpenAI, Anthropic, Google DeepMind, Meta and xAI, and by influential people outside them. They show who said a word and when, not how often everyone did.
- Rising, cooling and faded use one rule everywhere: the last 13 weeks (July 20 to October 18, 2026) against the 13 weeks before, and against the word’s best 13 weeks.
- Bubbles and blue chips are this site’s names for the two shapes a term takes on Hacker News: a bubble shot up within five months, a blue chip built over a year or more. The ticker shows each term’s last 13 weeks against the 13 before.
- Every line is drawn against its own peak, so charts compare timing and shape, not size.
- October 2026 is partial (1–18) and is left out of every comparison.
Hacker News: the crowd
For each week and month from October 2023 through October 18, 2026, the count is the number of distinct Hacker News accounts that posted a comment or story text containing the word. One person using a word ten times in a month counts once. The source is the public Hacker News dataset that ClickHouse keeps in sync with the live site. Story titles are not included, only text. Words with other everyday meanings count only when the same comment also mentions AI (a model, a lab, a chatbot or language models), and a few also leave out their known other uses: AGI in tax threads, hallucination next to psychedelics, multimodal next to transit, jailbreak next to phones.
“Per 1,000 active users” divides by every account that posted that month: about 45,500 in October 2023, dipping to 38,700 in late 2024 and rising to 54,900 by July 2026, so changes in the site itself do not read as hype. Weekly counts use the same patterns per Monday-to-Sunday week.
Status: a word is rising when its last 13 weeks run at least 25% above the 13 before and at least one other signal (Bluesky, Wikipedia, arXiv or GitHub) is also up; faded when its last 13 weeks are below 35% of its best 13 weeks; cooling when they are at least 20% below the 13 before; otherwise holding.
Wikipedia readers
Weekly page views of the English Wikipedia article closest to each word (for example “Vibe coding”, “AI slop”, “Model Context Protocol”, “Chatbot psychosis”), counted for people only, with automated traffic excluded, from the Wikimedia pageviews API. A line starts the week its article was created. Words with no article that enough people read have no Wikipedia line, and a few rest on broader articles (AI agents on “Intelligent agent”, interpretability on “Explainable artificial intelligence”, continual learning on “Incremental learning”, AI tells on Wikipedia’s own guide to signs of AI writing).
Bluesky posts
For each week, the number of Bluesky posts that use the word, from Bluesky’s public search (exact phrases, one or two spellings per word). Bluesky grew about tenfold in November 2024, so each week is divided by the number of posts using a fixed set of everyday words (umbrella, laundry, dentist, bicycle); the line shows the word’s share of conversation rather than the site’s growth. Bluesky’s search reports at most 10,000 matches per query and week. “Sandwich” and the bare word “slop” hit that ceiling, so they were left out: on Bluesky, slop counts only the phrase “AI slop” (on Hacker News, where the bare word is used almost only about AI, it counts “slop”). Single everyday words such as hallucination, humanoid and distillation also catch posts that have nothing to do with AI, so their level is less meaningful than their shape. Before November 2024 the site was small and weekly shares jump around more.
arXiv papers
New arXiv papers whose abstract uses the term, dated by first submission, from the arXiv API. Broad terms are narrowed to papers that also mention language models (for example reinforcement learning, distillation, hallucination, multimodal), and general terms to computer science. Weekly and monthly counts are divided by all new papers in machine learning, AI and computational linguistics in the same period. arXiv’s search matches word stems, so “agentic” would also find every paper about agents; agentic and interpretability are therefore counted by checking each matching abstract for the exact word. Seven words have no arXiv line because papers rarely use them, or use them for something else: AI slop, AI psychosis, AI bubble, AI winter, gigawatt compute, sovereign AI and evals.
Builders: installs, job posts, launches and GitHub
Job posts come from Hacker News’s monthly “Who is hiring?” thread (about 250–400 companies a month): the line is the share of posts using the word, for the ten words that appear in enough posts to count. Show HN launches are the stories whose title starts with “Show HN”, where people post things they built (about 1,000–1,500 a month in 2024, rising to nearly 4,000 by 2026 as coding agents made building cheap); the line is the share whose title uses the word, for the sixteen words that appear in at least 40 titles.
GitHub: for seventeen words that name something people build (AI agents, computer use, MCP, coding agents, multi-agent, deep research, vibe coding, context engineering, RAG, prompt engineering, evals, reasoning models, chain of thought, synthetic data, small language models, world models, long context), the number of new public GitHub repositories created each month whose name, description or topics use the term, from GitHub’s search API, divided by all public repositories created that month (about 3.5 million a month in late 2023, more than 11 million in September 2026). GitHub’s search matches word stems, so “agentic” would count every agent project and has no line.
Google Trends was also collected for the first 45 words and left out. In the first week of August 2025 nearly every technical AI search term jumped about threefold at once and kept climbing to as much as ten times its earlier level by spring 2026, then fell back over the summer, while everyday searches stayed flat. That looks like automated searches by AI tools rather than people.
Lab blog headlines
Lab people and outside voices: the record of statements
Outside voices are widely followed people who did not work at one of the five labs at the time: researchers, founders, investors, officials, journalists and writers, from Simon Willison, Ethan Mollick, swyx, Gary Marcus and Andrew Ng to the heads of other AI companies. Like the lab record, this is a curated sample of statements that drew attention, not a full count.
How the findings are computed
Limits
- Hacker News is a technical crowd. It over-represents developers and under-represents investors, policy people and the general public.
- The statements were collected with hindsight, by searching for words that later caught on. That can favor early mentions of words that caught on, which flatters “who moves first”. The last collection pass took every statement from each source it opened to reduce this.
- Matching is literal text. Words with other meanings count only in comments that also mention AI. That keeps out tax AGI, political sycophants and psychedelic hallucinations, but it can miss on-topic comments that never name AI, so those words read a little low.
- Lab headline counts read titles and summaries, not full posts, so they show what each company chose to headline.
- Downloads count installs by people and by automated builds alike, so they measure use, not people.
- Statements by lab staff and outside voices were collected for 45 of the terms. The rest are tracked on Hacker News, Bluesky, Wikipedia and lab blogs only.