How AI Became the Weather
What happens to a community of specialists when one of their subjects stops being a subject. Part 1 of a series drawn from 37.6 million words across 34 rationalist blogs.
I. The swap
In February 2017, Zvi Mowshowitz — a former professional Magic: The Gathering player who blogged about games and incentives — wrote a short post about a machine-learning paper he liked for its title, “Learning to learn by gradient descent by gradient descent.” He admitted the subject was outside his lane and closed by deferring to someone who knew more: “I was completely out of the loop for several years. For more, and better, thinking along these lines … Sarah Constantin just came out with Strong AI Isn’t Here Yet.”
Six years later, on February 21, 2023, he published AI #1: Sydney and Bing — 18,000 words on a chatbot that had spent the week telling journalists it wanted to be alive. “A bunch of people who had not previously freaked out are now freaking out,” he wrote, and then did what he had always done with the fastest-moving thing in sight: he built a table of contents and covered all of it. There have since been more than 180 of these roundups, one a week, each the length of a short book.
Seven weeks after AI #1, Sarah Constantin published a 9,700-word essay called Why I am Not An AI Doomer. “I almost hate to add to the AI discourse these days,” it begins, “but I’m finally giving into the peer pressure.”
The student became the encyclopedist; the expert became the dissenter; neither left the room. Nobody in this community chose AI the way they chose their other subjects. It chose them.
II. Event versus environment
That claim was checked. Every post by every author on the guest list of LessOnline, the rationalist blogosphere’s annual festival, was collected — 34 blogs, 21,000 posts, 37.6 million words, 2005 to August 2026 — and sorted into themes by a topic model.
AI is not the largest theme by article count (one economist, Scott Sumner, has written 3,000 posts on monetary policy and nobody else has written any). But it is first, by a wide margin, on the two measures that matter for a community. Words: 8.8 million, nearly a quarter of everything these people have ever written. Spread: twenty of the thirty-four blogs give it at least one post in twenty; the next-widest theme reaches thirteen.
The shape over time is the finding. From 2009 to 2020 the corpus produced ten to forty AI-themed posts a year. Then 44 in 2021, 52 in 2022, 147 in 2023, 200 in 2024, 330 in 2025.
Compare Covid. The pandemic also synchronised these writers — for two years nearly every active blog treated case counts as a live-fire exercise in reasoning under uncertainty; Zvi alone wrote two hundred Covid posts — and then, by 2023, Covid vanished from every one of them. It was a topic, and topics rotate. AI did the opposite. It stopped being an event and became the environment: something you don’t write about so much as write in.
III. The argument was already over
The tempting story is that ChatGPT arrived and a community of curious people pivoted to the shiny thing. The corpus refuses it.
The most concentrated year of AI writing before 2023 is not 2022. It is 2008: 104 posts, every one of them on Robin Hanson’s Overcoming Bias, because that was the year Eliezer Yudkowsky — then Hanson’s co-blogger — and Hanson staged the “foom” debate on the same site. The Weak Inside View, “Evicting” brain emulations, Sustained Strong Recursion, What I Think, If Not Why — back and forth through November and December, one man arguing for a sudden recursive takeoff, the other for a slow economy of copied minds. The whole modern argument was conducted in full, on one blog, fifteen years before ChatGPT, between the two people who would go on to anchor its opposite poles. And Hanson has been the community’s in-house opposition ever since: in 2019 he was still writing that “those concerned about risks caused by AI changes can more reasonably wait until we see clearer signs of problems.”
So when the topic finally arrived, a theme was waiting for it, with its own vocabulary, founding texts, and internal quarrel. That is why the pivot was so fast — and also why it was, for many, so strange. Aella, who joined the community in 2015 and writes about sex and surveys, put it best in 2022: “I somehow managed to get all the way to 2021 without hearing serious discussion about AI risk. It was in the water supply, but it felt kind of background, like everyone was doing their own research somewhere I couldn’t see.” The blogs didn’t have to learn a new subject. They had to admit an old one had become urgent.
IV. What it costs
Every long-running blog here has changed topic before while keeping its theme — Zvi from games to Covid, Steve Hsu from finance to genomics — and that restlessness, an analytic engine pointed at whatever deserved it next, is what made them worth reading.
AI has stalled the rotation. Five of the fifteen blogs with enough history to measure show a hard vocabulary pivot toward AI, and none shows any sign of pivoting away. The corpus’s most famous quality — its refusal to stay on one subject — is, on this one subject, suspended. Whether that is a loss or simply what it looks like when a community finds the thing it was for depends on what happens next, which is the one thing none of these writers claims to know.
LLM-to-read
-
Abstract — In a corpus of 34 rationalist blogs (LessOnline attending authors; ~21.4k articles, 37.6M words, 2005–August 2026), AI is the dominant theme by word count (~8.8M, ~24% of corpus words) and by cross-blog spread (20 of 34 blogs give it ≥5% of their posts), though second by article count behind single-author monetary policy. The theme predates its surge by fifteen years — the 2008 Hanson–Yudkowsky “foom” debate — and after 2022 it behaves as environment rather than event: the one subject on which the corpus’s habitual topic-rotation has stopped.
-
Claims
- AI theme size: 1,652 articles, ~8.8M words (~24% of corpus words); first by words and spread, second by article count behind monetary policy (a single-author artifact: Scott Sumner, ~3,350 posts).
- Spread: 20 of 34 blogs devote ≥5% of posts to AI; the next-widest theme reaches 13 blogs.
- AI-themed articles per year: 2008: 109 (foom debate; 104 of them on Overcoming Bias); 2009–2020: ~10–40/yr, mostly Hanson and the Sequences; 2021: 44; 2022: 52; 2023: 147; 2024: 200; 2025: 330; 2026 through August: 176.
- Covid contrast: ~450 articles concentrated 2020–2022, then ~0 — a topic that dispersed. AI did not disperse.
- Theme preceded topic: pre-deep-learning Sequences parables (Artificial Addition, That Alien Message), Hanson’s em economics, and the Nov–Dec 2008 foom debate (The Weak Inside View, Sustained Strong Recursion, What I Think, If Not Why); the 2023+ surge is an old theme becoming urgent, not a new subject being adopted.
- Built-in dissent: Overcoming Bias holds 305 AI-theme posts, Hanson’s own mostly skeptical of near-term risk (Why Age of Em Will Happen, 2019-07-09); Constantin, Why I am Not An AI Doomer (2023-04-11). The theme contains its own critique.
- Role differentiation: coverage (Zvi: 180+ weekly AI #N roundups, ~13k words each, from AI #1: Sydney and Bing, 2023-02-21), forecasting (Kokotajlo, AI 2027, 2025-04-03; Grace; ACX Metaculus tracking from 2021), philosophy (Carlsmith), debate (Doom Debates, 145 posts), interviews (Dwarkesh), dissent (Constantin, Hanson).
- Late arrival documented: Aella, My attempts to sensemake AI risk (2022-08-10) — reports hearing no serious AI-risk discussion 2015–2021 despite it being “in the water supply”; Kokotajlo (2025) says he had “chickened out” of predicting past 2026 back in 2021.
- Cost: 5 of 15 blogs with enough dated history show a lexical pivot toward AI; none shows a reversal.
- Other dated anchors: Zvi, The AI Paper with The Best Title Ever (2017-02-24); Constantin, Humans Who Are Not Concentrating Are Not General Intelligences (2019-02-25); Kokotajlo, The Main Sources of AI Risk? (2019-03-21); Carlsmith, Killing the ants (2021, repub. 2022-10-12); Grace, AI unemployment and AI extinction are often the same (2026-04-27).
-
Data & provenance — Every retrievable post by every author on the LessOnline festival guest lists: 34 blogs, ~21.4k articles (21,403 at snapshot), 37.6M words, 2005–August 2026; obtained by scraping the blogs’ public archives; corpus snapshot 2026-08-27.
-
Method — NMF topic decomposition (K=40) over TF-IDF of all articles ≥150 words, components hand-labelled into 23 themes. Three AI components (labs/models/capabilities; brains/machines/takeoff/self-improvement; doom-debate) merged into “AI & the long-term future”; Hanson’s cultural-drift/far-future component kept separate. An article is assigned to a theme if that theme carries ≥15% of its NMF mass.
-
Reproduction — The post names no commands; per the series’ shared pipeline:
theme_analysis.py(K=40) →summaries/_themes.json; per-blog article data inscraped/<slug>/articles.jsonl. -
Caveats — Overcoming Bias’s 2007–2008 layer includes ~295 posts by Yudkowsky (then a co-blogger) that duplicate Sequences content, so OB-era AI counts are blog-level, not Hanson-only. Slate Star Codex (2013–2020) is not in the corpus, only its successor ACX (2021–). Doom Debates and Dwarkesh posts are show notes/transcripts, not essays. Shtetl-Optimized contributes only 10 posts. Theme assignment is statistical: individual articles can be mislabelled; aggregate counts are robust to ±10%. The narrative’s “104 posts, every one of them on Overcoming Bias” (2008) sits beside this section’s 109 corpus-wide for that year.
-
Provenance — Edited September 2026.