One Tool, Eight Locks
A synthesis of themes 3–10 — epistemics, law, fiction, culture, medicine, physics, status, markets. Part 3 of a series drawn from 37.6 million words across 34 rationalist blogs.
In August 2007, the physicist Steve Hsu posted a spreadsheet tracing subprime-mortgage bonds through European banks — a year before Lehman fell. He held no credential in finance. He just kept the numbers and believed them.
That is the whole movement in one gesture. The corpus’s founding theme isn’t a subject; it’s a repair manual for belief, assembled by Robin Hanson in 2006’s Reasonable Disagreement: if two honest people share the facts, their disagreement is itself data, and someone owes an update. Everything else in this stratum of the corpus — the eight themes ranked third through tenth, roughly 7,500 articles between them — is that manual, pointed at a different lock.
David Friedman points it at the courts. Why is Tort Law like Accounting? (2006) is two sentences of pure epistemics in a lawyer’s coat: both fields round every probability to zero or one — 60% liability becomes a million-dollar debt, 40% becomes nothing — “because both depend on low quality decision-making mechanisms.” He has written 327 such posts.
Hanson points the same instrument at himself, in near-total monopoly: 697 of 701 posts in the status theme — 99.4% — on why we perform status rather than merely seek it, from Why Pretend? in 2010 to Status Leadership in 2023. That is epistemics turned on deception rather than error, and it is the most concentrated theme in the taxonomy after monetary policy.
Yudkowsky dramatises the manual as 122 chapters of Harry Potter, and Alicorn’s Luminosity adds 436 articles more; between them they are the largest single word-count contributors to any theme in the corpus, and fiction runs to 4.77 million words. The point is never entertainment — narrative is an argument-testing device. The same apparatus even grades books: ACX’s annual anonymous book-review contest applies the evidence-and-update machinery to literary judgment.
Hsu works two of the locks himself, at different rigour: 60% of the physics theme is his alone — interviews, papers, academic-adjacent posts — and, with Overcoming Bias and TheMoneyIllusion, he anchors the markets theme his 2007 subprime call came from. And Hanson tested the manual against real medicine for thirteen years, review by review, and got his randomised trial in the end — part 2 of this series tells that story in full.
Add it up and the pattern inverts part 1. AI is the theme twenty blogs share; these eight are the themes one writer at a time carried into a domain and kept. Status is 99.4% one blog. Physics is 60% one author. Law is Friedman, 327 posts deep. In this stratum the rule is concentration, not convergence: each theme lives in whichever writer carried the habit into that domain first.
Eight themes, one habit, never turned back off.
Corpus and method as in parts 1 and 2: every post by the authors on the LessOnline guest lists — 34 blogs, 21,403 articles, 37.6 million words — sorted into 23 themes by topic model.
LLM-to-read
-
Abstract — Themes ranked 3–10 in the LessOnline corpus theme analysis (epistemics; law/policy; fiction; books/culture; empirical studies/medicine; physics; status/norms; markets/prediction — 7,543 articles, 15.8M words combined, ~40% of the labelled corpus) share a common structural pattern: each is the corpus’s founding epistemic method (evidence-weighting, disagreement-as-data, probability-talk) applied to a different subject domain, largely by a small set of authors reusing one analytical habit across unrelated fields. Against the AI theme’s 20-blog spread (part 1), concentration in single authors is this stratum’s rule.
-
Claims
-
Combined size: 7,543 articles, 15.8M words, ~40% of the labelled corpus (see caveats).
-
Per-theme statistics:
Theme Rank Articles Words Blogs ≥5% Dominant contributors Epistemics, argument & belief 3 1,215 1.70M 13 Overcoming Bias 656; read-the-sequences 210; David Friedman 104 Law, policy & governance 4 1,158 1.56M 7 David Friedman 327; Overcoming Bias 317; TheMoneyIllusion 293 Fiction 5 1,108 4.77M 12 Luminosity 436; HPMOR 121; Alexander Wales 111 Books, films & culture 6 921 2.31M 10 David Friedman 211; Overcoming Bias 196; ACX 99 Empirical studies & medical evidence 7 898 2.10M 7 Overcoming Bias 400; ACX 134; gwern 84 Physics & science 8 851 1.02M 6 infoproc/Hsu 508 of 851 (60%); Overcoming Bias 212 Status, norms & social behaviour 9 701 445K 1 Overcoming Bias/Hanson 697 of 701 (99.4%) Markets, finance & prediction 10 681 932K 4 Overcoming Bias 218; infoproc 208; TheMoneyIllusion 108 -
Epistemics is the method, not a theme among themes. Founding text: Hanson, Reasonable Disagreement (2006-12-06, OB) — persistent disagreement between honest truth-seekers is itself evidence one side hasn’t updated; the norm (disagreement-as-data, probability over certainty) recurs as vocabulary and structure across every other theme in the set.
-
Status is epistemics turned on deception rather than error: Hanson, Why Pretend? (2010-04-24) and Status Leadership (2023-07-13); near-total single-blog ownership (99.4%) makes it the most concentrated theme in the 23-theme taxonomy after monetary policy.
-
Law becomes a dataset: Friedman, Why is Tort Law like Accounting? (2006-02-11, 372 words) — both fields round continuous probabilities to binary liability (p<0.5→0, p>0.5→1), attributed to “low quality decision-making mechanisms”; the economics-of-law method applied at essay length, repeated 327 times.
-
Fiction is the method dramatised: HPMOR (Yudkowsky, 122 chapters) and Luminosity (Alicorn, 436 of the theme’s articles) are the corpus’s largest single word-count contributors to any theme (4.77M words) — narrative used as an argument-testing device, not entertainment. ACX’s annual anonymous book-review contest (e.g. Your Book Review: Breakdown In Pakistan, 2026-07-24, 7,214 words) applies the same evidence-and-update apparatus to literary judgment.
-
Markets: an early, verifiable call — Hsu (infoproc), Profiting from the meltdown (2007-08-10, 1,204 words) tracks subprime ABX tranche pricing and European bank exposure over a year before the September 2008 Lehman collapse, followed by real-time crisis coverage (Mark to market, 2007-06-21; AIG watch, 2008-11-10) — the corpus’s clearest evidence-tracking-precedes-events example.
-
Physics and medicine are the same author’s two hobbies at different rigour: Hsu supplies 60% of the physics theme (interviews, papers, academic-adjacent posts) with only light connection to his genomics main line; medicine is Hanson’s decade-plus campaign (2007 Medical Study Biases through 2010 Supplements Kill to the 2021 Karnataka RCT he’d argued for) — full treatment in part 2, briefly reprised here to complete the pattern.
-
Concentration is the rule, not the exception, in this stratum: every theme in 3–10 except epistemics and fiction has one author supplying ≥40% of its articles (status 99.4%, physics 60%, law 28%, markets 32%, culture 23%); contrast the AI theme (part 1), spread across 20 blogs — themes 3–10 are where individual voices, not communal convergence, dominate.
-
-
Data & provenance — Every retrievable post by every author on the LessOnline festival guest lists: 34 blogs, 21,403 articles, 37.6M words, 2005–August 2026; obtained by scraping the blogs’ public archives; corpus snapshot 2026-08-27.
-
Method — Same NMF/TF-IDF pipeline as parts 1 and 2 (
theme_analysis.py, K=40 components, 23 hand-labelled themes, outputs insummaries/_themes.json). This post treats themes 3–10 (by article-count rank) as a set and extracts the cross-theme pattern rather than analysing each in isolation; sources pulled directly fromscraped/<slug>/articles.jsonl. -
Reproduction —
theme_analysis.py→summaries/_themes.jsonfor counts, ranks, and spread; per-blog article data inscraped/<slug>/articles.jsonl. -
Caveats — Theme boundaries are NMF-component merges, hand-labelled; some articles are misclassified at the margin (share <0.3 near cluster boundaries). “Books, films & culture” and “Fiction” overlap partially by construction (both draw on narrative-adjacent vocabulary) — treated as distinct themes per the original taxonomy, not re-merged here. Overcoming Bias’s 2007–08 layer includes ~295 Yudkowsky-authored posts (co-blogger era) that also appear in read-the-sequences; this affects epistemics-theme attribution at the margin. Word/article counts drawn from
summaries/_themes.json, corpus snapshot 2026-08-27 (21,403 articles, 37.6M words). The headline combined size (7,543 articles, 15.8M words) does not match the sum of the per-theme table (7,533 articles, ~14.8M words), and the “≥40% single-author” claim is contradicted by three of its own parenthetical figures (law 28%, markets 32%, culture 23%). -
Provenance — Edited September 2026.