The room where the arguing happens

September 2026 · ILIAD

One of three essays from a single corpus study of 49 theoretical-alignment (“ILIAD”) authors: this one reads the social corpus (3,653 posts, Aug 2025–Aug 2026); its companions read the long-form archive (2,743 pieces, 2003–2026) and the cross-corpus synthesis.

There is a particular sound a field makes when it is talking to itself, and if you stand inside 3,653 posts written by forty-nine theoretical alignment researchers between August 2025 and August 2026, you can hear it clearly. It is not the sound of pronouncement. It is the sound of qualification — the endless, scrupulous, faintly exhausting noise of people trying very hard not to overclaim.

The first thing that strikes you is how little of this is writing at all, in the sense of a person sitting down to say a thing. Nearly a third of these posts — 30.8% — carry no topical vocabulary whatsoever. They are agreement, thanks, jokes, logistics, a raised eyebrow. “They’re the same picture.” “Huh? Example?” “Fair enough!” Roughly a fifth are under a hundred characters. This is not a corpus of essays with occasional replies; it is a corpus of replies with occasional essays, and the essays sit on top of the conversational mass like a thin crust of rock on a great deal of water.

And then there is the thing that isn’t there.

This community has a reputation — inside and outside itself — as the place where people put numbers on their beliefs. P(doom). Credences. Calibration. Three independent readers went through three independent 244-post samples looking for it. They found the phrase “epistemic status” exactly once, once, and zero times. Numeric probabilities of any kind appeared in roughly 4–7% of posts — and several of those turned out to be parameter counts rather than beliefs; genuine first-person probabilities about the world ran closer to 2–4%. Meanwhile “I think” showed up in about a fifth of everything, “seems” in another seventh, and parenthetical self-qualification — the sixty-character aside that walks back the sentence it is attached to — in 22%.

The calibration culture is real. It has simply been absorbed into grammar. These people no longer perform Bayesianism; they decline nouns with it. Paul Rapoport answering a three-part question entirely in graded negations — weak no, no I think, wouldn’t be surprised if it were a weak no — is not hedging. That is the load-bearing structure of the sentence.

Characters

A corpus of this kind is really a cast, and the cast sorts itself with unusual clarity.

John Wentworth never posts on X. Not once, across three samples. He writes only on LessWrong, at length, in a register so flat and so consistent that it does not change when the subject does. He applies the identical prose to Solomonoff induction, to double-entry bookkeeping, to fashion semiotics, to his own genome. And that last one is the thing you cannot look away from: a post working out that he does not feel companionate love, pursued as a forensic investigation — self-sequencing, a single-base-pair deletion in the oxytocin receptor gene, an honest note about which conclusions short-read sequencing cannot support. Confessional autobiography with a methods section. Nothing else in the corpus does this, and I am not sure anything else anywhere does.

Eliezer Yudkowsky is the corpus’s range. He holds both extremes at once: a four-word tweet and a 29,000-character defence of being unusually confident; a public request that people falsify his claims about OpenPhil before he publishes them; a genuinely beautiful eulogy for his grandfather that opens by refuting a proverb about death on semantic grounds before grieving. He is also the only reliable source of contempt in a corpus that otherwise cannot bring itself to be rude, and the only one whose insults are built out of technical vocabulary — an opponent reasoning “at GPT-2 level,” a country reduced to “a scary word-vector.”

Mikhail Samin is the activist, and the only voice that pleads. After a Molotov cocktail was thrown at Sam Altman’s house he wrote a flat anti-violence post that ends by saying he wants absolutely everyone to survive, Altman included. He is also the corpus’s most willing litigant — quoting DMs, disputing word counts, telling CEA’s leadership to resign in a single sentence with no hedge at all.

Emmett Shear marks the boundary. He writes on X only, in a gnomic process-philosophy idiom with no citations, no probabilities, no hedges — covariance and Heyting algebras and alignment as a direction rather than a destination. Read him next to Vanessa Kosoy’s numbered, declarative, textbook-cadenced formalism and you are looking at the outer edge of what this community can absorb.

Jörn Stöhler is the most interesting structural case: 100% on X, and yet writing pure LessWrong — six-point reconstructions of an opponent’s model, raw odds notation in a tweet, a dated 16% forecast with a mechanism list. Which settles a question the corpus keeps raising. Platform predicts length. Author predicts voice. X strips the blockquote, the footnote, the acknowledgements and the ability to edit; it does not change who anyone is. (One caveat, disclosed in the data notes below: the X handle behind these posts is attributed to Stöhler but not confirmed.)

What moves the needle

Across thirteen months the themes are more stable than you would expect from a field that describes itself as moving fast. Deep learning and LLM capabilities hold 16% throughout. But two moments show through the noise. November 2025 spikes hard on research practice (20.9%), rationality (21.2%) and personal writing (11.0%) — that is Inkhaven, a thirty-day daily-blogging residency, visible in the data as a month where a whole community wrote more and wrote worse and said so in the text: written quickly, apologies, written in the shower. And the summer of 2026 shows governance climbing while conversational chatter climbs with it — the signature of a field arguing about institutions rather than theorems.

The deeper finding is what the arguing is made of. Not evidence, mostly. Analogy. The economy as a mean-field system; gas particles behind a sliding door for AI program-space; a trained immortal dog implementing a language model; heroin as pre-existing value-overwrite; a retired engineer’s chalk mark. When these people want to establish something, they reach for a mechanism from a distant domain and carry it over. It is the community’s default proof strategy, and it is used in technical arguments, not just rhetorical ones.

The thing worth admiring, and the thing worth worrying about

The admirable thing is the retraction norm, and it is not a pose. About 3% of posts contain a visible concession, correction or public reversal — and crucially, the correction is left in place, as a scar. Strikethroughs preserved. “EDIT:” markers that disown half a post. Richard Ngo publicly rewriting “almost the entire field” down to “most of the field” and leaving both visible. In most professional discourse the edit is a silent tidy-up. Here it is treated as evidence, and destroying it would be destroying data. This is genuinely unusual among professional corpora, and the community deserves the credit it gives itself.

The worrying thing is adjacent to it. The same flat, analytical register that makes the self-correction possible is applied, without modulation, to other people. To a colleague’s sexuality. To a stranger’s weight, recast as a self-honesty failure. To women, in cost-benefit terms. The corpus contains a small but real vein of writing where the method is unchanged and the object of study is a person who did not consent to be one. That is not a failure of rigour. It is rigour pointed somewhere it does not belong — concentrated, the counts say, in one author, but consistent enough across thirteen months that it cannot be read as a one-off.

My opinion

I think this is a community that has solved a hard problem — how to argue in public without lying about your confidence — and has mistaken that solution for a general-purpose method.

The epistemic machinery is real and it works. The hedging is not cowardice; it is precision, and the near-total absence of numeric credences alongside near-total presence of verbal ones is, I would argue, correct behaviour rather than hypocrisy. Numbers where you have no model are theatre. These people mostly decline to perform it, whatever their reputation says.

But the machinery is load-bearing in places it cannot bear. The most-quoted register in this corpus — reduce the claim to a mechanism, deny the mechanism — is superb against a bad argument and useless against a person. And the corpus’s own most consequential fact goes almost unexamined: Lucius Bushnaq mentioning, in passing, in a shortform, that his mathematics research is now substantially done by AI agents working in parallel, and that the bottleneck is that they explain themselves poorly. That is the field’s subject matter arriving inside the field’s daily practice, and it is recorded the way one records the weather.

If I were a researcher here, I would take the calibration norms as settled and spend my worry entirely on that last point.

LLM-to-read

Abstract

Analysis of the social corpus of a study of 49 theoretical-alignment (“ILIAD”) authors: 3,653 cleaned posts (5.78M chars) from X, LessWrong/Alignment Forum and the EA Forum, 2025-08-29 to 2026-08-29. Half the corpus is conversational register; the community’s reputed numeric-calibration practice appears as verbal hedging rather than explicit credences. Platform predicts post length while author predicts voice; hedging intensity is predicted by claim-checkability. A visible-retraction norm (~3% of posts, corrections left in place) is documented, as is a small vein of writing that applies the community’s flat analytical register to non-consenting third parties. Monthly theme shares are stable across the 13-month window apart from a November 2025 residency spike and rising governance discussion through 2026.

Claims

Primary-theme distribution (share of 3,653 posts; 17 themes, bottom-up lexicons, 100% coverage):

• 1,805 of 3,653 posts (49.4%) contain zero technical-lexicon terms; every post that scores at all clears the threshold comfortably, so this is not a threshold artefact. Half the corpus is conversational register.
• The six themes the community is named for (Agent Foundations, Mech Interp, SLT, Natural Abstraction, Computational Mechanics, Pure Maths) total 10.4% of social posts.

Time-series (monthly, share of posts touching each theme): stable across the 13-month window; two signals exceed noise:

• November 2025: Research Practice 20.9% (vs 13.5% mean), Rationality 21.2%, Personal 11.0%. Cause identifiable in-text: the Inkhaven daily-blogging residency. Accompanied by explicit low-effort disclaimers (“written quickly”).
• 2026 H1–H2: AI Governance rises to 10–13% in Feb/Mar/Jul 2026 from 5–7% in autumn 2025.
• Deep Learning/LLMs holds 15–27% every month; no trend.
• Conversational share is volatile (14.4%–51.5%) and inversely tracks long-form months.

Common features (from full reading of a 20% theme-stratified sample, n=732, 3 independent readers):

1. The reply is the unit, not the post. ~21% of documents are antecedent-free fragments unintelligible without their parent.
2. Blockquote-and-answer is the LessWrong genre marker. ~14% of all docs; ~20–25% of LW comments open with > and walk down the parent point by point. Opponents are pasted, not summarised.
3. “Epistemic status” is effectively absent: 1, 1 and 0 occurrences across three 244-doc samples. Where present, it is used ironically.
4. Numeric credences are rare: ~4–7% of docs contain any probability, percentage or credence; genuine first-person probability estimates about the world are ~2–4%. Verbal hedges outnumber numbers roughly 10:1.
5. Hedging is scoped, not global: a confident assertion plus a narrowly delimited uncertainty. Authors hedge the inference, not the assertion.
6. Parenthetical self-qualification (60+ char aside) in ~22% of docs — the dominant hedging organ.
7. Public self-correction is a norm: ~3% of docs contain a visible retraction, concession or EDIT: marker, left in place rather than silently revised.
8. Analogy is the primary proof strategy, including in technical arguments — mechanisms imported from physics, biology, economics, engineering.
9. Disagreement protocol: concede first, locate the crux as definitional, then counterattack. “Agree/agreed” outnumbers “disagree” ~2.7:1 in an overwhelmingly argumentative corpus.
10. Humour is deadpan and marker-free: emoji or :) in ~3–4.5%, profanity ~1%.

Patterns in the differences:

• Platform predicts length; author predicts voice. Medians: X 159–193 chars, LW comment 610–830, LW post 5,734–10,021. Authors who post on both barely modulate register. X removes blockquoting, footnotes, acknowledgements and editing — not style.
• Author platform-choice is bimodal. X-only: Stöhler, Shear, Rosas, Hobbhahn, Sandberg, Shai, Murfet. LW-only: Wentworth (0 X posts across all three samples), Hänni, Demski, Kosoy, Rapoport, Ngo (mostly). Yudkowsky is the inverse of Wentworth.
• Hedging is predicted by checkability, not by platform or author. Technical claims are asserted; community-strategy and forecasting claims are heavily hedged.
• Length as author fingerprint: Rosas ~67 chars median; Yudkowsky ~192; vs Wentworth ~983–2,455, Hänni ~3,072.
• Theme concentration by author: mathematical themes are almost entirely five authors (Wentworth, Hänni, Demski, Lorell, Kosoy) on LessWrong.

Outliers (highest-signal, verified across readers):

• Register outlier class worth flagging: a small but consistent vein (~1–3% of docs, concentrated in one author) applies the corpus’s flat analytical register to non-consenting third parties — colleagues’ sexuality, a stranger’s body, women in cost-benefit terms. Method unchanged; object inappropriate.

Assessments (the post’s argued positions):

1. The community’s reputation for numeric calibration is not supported by its social output. Calibration is present as syntax, not arithmetic. Assessed as appropriate rather than hypocritical: numeric credences without models are noise, and these authors mostly decline to produce them.
2. The public-retraction-left-visible norm is the corpus’s strongest and most transferable practice. It is rare in professional discourse and should be read as a genuine methodological achievement.
3. The dominant argumentative move (reduce claim to mechanism, deny mechanism) generalises badly from arguments to people, and the corpus contains visible casualties of that generalisation.
4. Highest-value under-examined datum: a working researcher reporting, in passing, that AI agents now do the bulk of his mathematics. The field’s subject matter has entered its daily practice and the corpus has not metabolised this.

Data & provenance

• Source: social-media.json, 3,653 posts after cleaning (5.78M chars), 49 authors, 2025-08-29 to 2026-08-29.
• Platforms: X 2,048 posts; LessWrong/Alignment Forum 1,851; EA Forum 18. Bluesky yielded nothing in-window (only one author has an account; last post March 2025).
• Acquisition: X retrieved via a date-bounded commercial search API; forum posts scraped per the study pipeline.
• Cleaning: URLs, HTML, CDN/CSS artefacts, LaTeX commands and markdown image refs stripped before analysis.

Method

17 themes derived bottom-up from unigram and bigram frequency, then formalised as weighted phrase lexicons; every document scored length-normalised and assigned a primary theme (coverage 100%, 0 unassigned). Monthly time-series over the 13-month window. A 20% theme-stratified sample (n=732) was read in full by 3 independent readers, whose converging observations supply the qualitative findings.

Reproduction

• Cleaning: tools/corpus.py. Full pipeline in the study’s tools/ directory (documented in the synthesis essay).
• Taxonomy: themes-topics.txt.
• Corpus file: social-media.json (includes the in-file flag on the unconfirmed X handle).

Caveats

The post’s own:

• X data retrieved via a commercial search API that is date-bounded rather than count-bounded; high-volume accounts are sampled less deeply than quiet ones.
• Five X accounts returned zero posts on cheap calls (DeDeo, Garrabrant, Goldstein, Heimersheim, Riechers). Treat as unconfirmed, not as verified silence.
• One X handle (@JStoehler, attributed to Jörn Stöhler) is not confirmed to belong to that person; 214 posts are included but flagged in social-media.json.
• Long posts are truncated in analyst bundles; body-text counts are lower bounds.
• Theme labels are automatic and noisy at roughly the 5–10% level; treat any single label as weak evidence.

Editorial:

• The human section glosses the 30.8% conversational-primary share as posts carrying “no topical vocabulary”; the zero-technical-term share is the separate 49.4% figure above. Both kept as the author wrote them.
• The named-author “Characters” section characterises writing style from sampled public posts. Consent questions around quoted semi-private individuals, and around the unconfirmed-handle attribution, remain open.

Provenance: edited September 2026.