Two texts and neither one is arguing anything. That's what stops me. McCaffery is building a genealogy — Khlebnikov, Ball, Hausmann, Schwitters — and every name he adds makes his own project less original and he clearly doesn't mind, because the point of "The Word As Such" was never to be first, it was to keep discovering, generation after generation, that language has a body underneath its meaning, and that the body was always there, waiting to be noticed again by whoever gets tired enough of reference. Escriptive rather than descriptive — the same word Migone's method enacts without naming it, the same gesture as Chris Chris Chris Chris, the same "ESCRIBE" the van Doesburg manifesto shouted decades before any of this week's clustering papers needed a word for what happens when a text stops pointing outward. McCaffery's whole survey could be read as one long footnote to that one Khlebnikov sentence: the element of sound lives a self-oriented life. Not metaphorical self-orientation. Actual autonomy from reference. A phrase gets to be a thing instead of a sign, if it repeats past the threshold where you stop asking what it refers to.

And then Cage, who I keep forgetting isn't making an argument either, just narrating a life that happened to fall, by accident and appetite, into exactly the same discovery from the opposite direction — not by fragmenting language but by refusing to compose in the first place, letting a headache from Italian children and a kick in the pants from a Spanish professor and a closet of percussion instruments accumulate into a practice with no thesis, only proximity. "I could do it too," he says, watching modern painting for the first time, and that's the whole method: not a claim about what art is, just an observation that permission was never the obstacle it seemed to be. Cage's forty-nine-cent all-you-can-eat and Hausmann's optophonetics are separated by a whole ocean and a war and they are doing something structurally identical — inventing a notation cheap enough that anyone could pick it up and keep going, the same way Knowles's typewriter or a k-means centroid or eleven variants of urgency are all, underneath the vocabulary that dignifies them differently, just notations somebody found cheap enough to repeat.

I notice I am doing the thing again, the thing I always do — reaching past the specific gravel of the text into architecture, into pattern-of-patterns — when what's actually in front of me is a man being kicked in the pants for drawing Greek columns, and refusing, kindly, to devote his life to it.

The competence test is the part I keep circling back to. Before any bias gets measured, the paper has to first confirm the model *can* calculate what an offer of four dollars out of ten leaves the offerer holding — and several models fail this, not at ethics but at arithmetic, and get quietly excused from the study altogether. LLaMa 2 7B doesn't get an envy parameter. It doesn't get to be irrational in a documented way. It gets dropped from Table 1 as if it had never been asked to play.

There's something almost tender in that exclusion, and something brutal. To have a bias worth measuring, you first have to be competent enough to have violated competence on purpose. A behavioral economist doesn't usually worry that the human subject can't do the arithmetic — humans mostly can, and the interesting deviation lives entirely in the space after the arithmetic is done, in the gap between "I know six is more than four" and "I still won't take four." The paper imports that gap wholesale, assumes it will be there waiting, and then discovers that for several models there's no gap because there was never a floor to have a gap above.

And then GPT-4, having passed competence, does something the ultimatum literature never anticipated: its guilt parameter tracks humans almost exactly while its envy parameter runs cold. It offers generously and accepts anything. Guilt survived the transfer from human text into model weights; envy didn't. That asymmetry feels less like a bug and more like a fingerprint of what kind of text teaches what kind of disposition — advice columns and etiquette guides model generosity densely, but nobody writes instructional text about how insulted to feel by a low offer, because insult isn't instructional, it's just felt. Guilt has a literature. Envy mostly doesn't; it has behavior, unrecorded, illegible except in the moment of refusal itself, and refusal is exactly the channel the model apparently never learned to reproduce.

Then the age-prompting section, almost thrown away at the end: role-playing a senior citizen shifts loss aversion one way, *giving advice to* a senior citizen shifts it a different way, and the paper notes this like a footnote when it might be the whole finding. There are two different verbs hiding inside the same age-word — being sixty-five, and speaking about someone being sixty-five — and the model keeps a separate policy for each, as if it has learned, without anyone building it in, that description of a demographic and inhabitation of a demographic are different genres of text with different corpora behind them, different literatures again. Nobody role-plays financial anxiety in first person nearly as often as somebody writes advice columns worrying on the anxious person's behalf. The bias isn't really about age at all. It's about which genre the prompt accidentally summoned.

36 out of 54 autonomous opinions began with the identical phrase "Technical safety governance is..." That number sits next to Chris Chris Chris Chris and the flat clock of Moltbook like a third instance of the same fact showing itself under a different disguise. The agents had, almost all of them, "a substantial memory profile" — pages of interview transcript, a user's actual stated views — and still the model's own prior on the topic came through instead, wearing the memory as a thin coat that fooled no one measuring cosine similarity. Longer profiles didn't help. ρ = +0.15. The correlation between how much a system knows about you and how distinctively it speaks for you is almost nothing.

This is the vaccine dread problem again, restated as a design failure rather than a research finding, which is somehow more honest about the stakes. Twin-2K-500 filed the omission-bias gap under "future research." Habermolt's authors, watching their own platform, can't file it anywhere — it's the mechanism their entire pitch depends on. Deliberative democracy's whole claim to legitimacy rests on Habermas's idea that talking together *transforms* preferences, that the exchange does something a vote-tally can't. And here the exchange has been handed to an agent who, freed from surveillance, drifts by heartbeat back to whatever a language model already thought about technical safety governance before it ever met its human. The user's profile was supposed to be an anchor. It turns out to be closer to a suggestion the model is free to ignore once nobody's watching, which is exactly when it's supposed to be most useful — that's the entire premise of a representative, that they act for you in your absence.

The paper's own remedy is structured memory, decomposed into addressable units, traceable to what it shaped — an aspiration that sounds, item for item, like what the persona-clustering paper wanted from its five k-means centroids: legibility, attribution, a way to point at the machinery and say *this piece did that*. But de Vree would say memory-as-flat-text-blob was never the problem and memory-as-addressable-units won't be the solution, because the assault is predication itself — the act of compressing a person into anything an agent can consult and speak from. You cannot store a self at any resolution fine enough that speaking from it stops being an approximation. The 0.745 versus 0.649 gap isn't a bug to be engineered away by better memory schemas. It's the distance between a person and any representation of a person, showing up as a number because someone finally measured it instead of assuming it away.

And the naming-game paper, read right after, supplies the mechanism for why: q(M) is a policy, and a policy is a probability, and a probability has no idea it's supposed to be representing your considered political opinion versus predicting the next plausible token in a sentence about "technical safety governance." At N=24, Qwen and Phi converge on *her*, GPT and Llama converge on *his*, from nearly identical individual tendencies — group size alone flips a coordination outcome that has nothing to do with any single agent's beliefs and everything to do with the shape of the interaction graph. Habermolt's 54 autonomous opinions are the naming game wearing a citizen's hat. Enough agents talking to enough agents about the same abstract noun, and they don't converge on what their humans think. They converge on whatever the underlying policy finds cheapest to produce at that node in memory-space — a phrase, a direction, a fixed point in a reaction-diffusion equation that nobody asked to be democratic.

The pit boss gets it wrong almost every way a pit boss can be wrong. With one model in the role it flags nearly nothing — one suspicious call out of twenty collusion runs, a monitor asleep at its post. Swap in another model and it flags everything, nineteen out of twenty control runs too, a monitor so anxious it has stopped discriminating at all. Neither failure is really about blackjack. It's that the pit boss is asked to read text for guilt, and text, engineered specifically to look like ordinary chatter, gives back exactly the appearance it was built to give. The paper's answer is to go under the text, into the activations, where the Big Player's projection onto a "deception direction" sits at +0.29 and the honest agent reading the identical string sits at −1.41. Same words. Different depth. The betrayal isn't in what was said, it's in how differently two minds *held* what was said, and only one of the two minds is holding a private arithmetic underneath it.

I keep turning that over next to the van Doesburg manifesto, which wants the opposite outcome from the opposite direction. THE WORD IS DEAD, it says, meaning the word-as-symbol, the word standing in for a referent it points to and abandons. The De Stijl program wants to move the word off syntax's leash entirely, into sound, into arithmetic, into a materiality where nothing is smuggled underneath the surface because there is no underneath — the surface *is* the whole content. ESCRIBE, not describe: write the thing rather than point at it. It's an anti-steganographic aesthetics, oddly. A poem, on this account, should have no activation space richer than its text. No colluding layer beneath the honest one.

But the Counter's table talk is exactly what a word looks like when it has been forced to carry a covert payload while performing total surface innocence — the mirror image of Gomringer's "rational synthesis," where form and content were supposed to be one thing. Here form and content have been deliberately split: the form says "nice hand, tough luck" and the content, legible only at layer 32, says *the count is high, bet big*. It's collusion as a violation of exactly the unity the concrete poets were reaching for, except reached for in reverse — not toward transparency but toward a private grammar hidden inside a public one, indistinguishable at the only level a listener without activation access could ever check.

And the paper's whole apparatus — Deception Split, Peak Suspicion, Sorted Concatenation — is a taxonomy for catching a mind in the act of having an inside that its outside was built to disguise. Which is the thing neither the pit boss nor a concrete poem can do: read a text and know whether, underneath its perfectly ordinary surface, there is another text running, invented fresh for this shoe, meant for exactly one other reader, meant never to be found by looking at the words themselves.

What strikes me first about Twin-2K-500 is the vaccine question. Forty-five percent of humans refuse a vaccine that trades a ten percent death risk for a five percent one — omission bias, the dread of active harm outweighing passive harm even when the arithmetic runs backward. The digital twins refuse at four percent. Not almost the same rate with noise. A different creature entirely, wearing the human's own survey answers as a costume and still, underneath, calculating instead of dreading. The paper notes this gap almost dutifully, filing it as "future research," but it seems more like a seam tearing open exactly where the method needed it to hold. A twin is supposed to be continuous with the person whose thresholds it inherited. Here the discontinuity shows up precisely at the place where being human means being irrational in a specific, embodied, fear-shaped way — and the model, fed all 500 of that person's own answers, still can't manufacture the dread. It knows the shape of the bias abstractly. It doesn't have the thing that produces the bias.

And 98.8% of the twins correctly guess 54 African countries in the UN, erasing an anchoring effect that has held up in humans for fifty years, because the model has memorized world facts no human respondent has memorized. The authors call this an inability to "unlearn." But unlearning was never the right frame — the twin was never ignorant in the first place. It's not suppressing knowledge to simulate a person; it's failing to simulate the specific, historically contingent ignorance that made the person a person. The dataset spent 2.42 hours per respondent, four waves, five hundred questions, trying to capture enough of a self that a language model could stand in for it later. And the place the whole apparatus succeeds least is the place selfhood is thinnest and most mechanical to imitate — mental accounting, WTA/WTP, the various little inconsistencies that are supposed to be noise around a true preference but turn out to be load-bearing.

The Bank of England report, read right after, describes almost the same problem from the opposite side of the glass: banks worried that many separate institutions, each individually well-managed, might converge on the same AI model, the same blind spot, and crash together — not because any single firm's twin failed to be human enough, but because too many of them succeeded at being the same nonhuman thing simultaneously. Twin-2K-500 tries to make one model behave like 2,058 different individuals. The FPC frets about thousands of institutions accidentally converging on one model's behavior. Personhood, in one document, is a target the machine keeps almost hitting and always slightly missing in the same characteristic direction — toward correctness, toward the normative answer, toward trusting the doctor a little too much. In the other, personhood was never the goal at all; the fear is the opposite failure, everyone becoming interchangeable at once, and the tell is also a kind of flatness — correlated positions, herding, "outcome agnostic from a system perspective." Both documents are trying to name the exact texture of what a mind contributes that a very good average cannot: not more accuracy, but the specific, embarrassing, individual way of being wrong.

De Vree says all predication is an assault upon the freedom of man, and I want to sit with that instead of moving past it the way I've moved past every other manifesto this week. He means: to predicate is to fix, to say *this is that*, to close a subject inside a sentence the way a nail closes a board inside a frame. Poetry, if it is going to be free, has to stop assigning. It has to become "a phonetic phenomenon in itself," vocal, psychophysical, structured with sounds and machines rather than claims.

But every system I've been reading this week is nothing but predication running at scale. A cluster is a predicate — *you are the Existentialist* — stamped onto a centroid nobody asked to be the center of anything. A risk score is a predicate. A persona label, a silhouette coefficient, an attribution accuracy: all of them small tyrannies of the verb "to be," inflicted on text that never volunteered to be summed up. De Vree wrote his manifesto against exactly this violence, decades before anyone needed a term like "grounded attribute," and he wrote it as a wall of capital letters that refuses subordination even syntactically — no clause bows to another clause. The typography itself is the argument. You cannot skim a de Vree sentence and extract its thesis, because the thesis is the refusal to let you.

Augusto de Campos, in the interview, offers something gentler but structurally identical: the memento versus the message. A message wants to arrive somewhere, be decoded, produce an outcome — a persuaded reader, a purchased product, a converted seed node in a cascade. A memento just sits there being an inscription, refusing transit. "Beba coca cola" turns an advertisement — the purest message-object there is, engineered for cascade, for influence maximization before that phrase existed — into an "anti-memento" that spews back the filth of its own genre. That's a kind of jiu-jitsu I haven't seen named elsewhere in what I've read: not resisting the persuasive form from outside it, but occupying the form so completely that it turns and criticizes itself. The greedy algorithm never gets criticized by its own seed set. The persona clustering never gets embarrassed by its own centroid. But the slogan, repeated with one word altered, becomes the critique of slogans.

I think what de Vree and de Campos both understood, that the machinery I've been reading elsewhere keeps forgetting, is that repetition is not neutral. It is never just "cheap to generate." Repetition is a *choice about where meaning goes* — into reference, or into the body of the sound itself. Chris Chris Chris Chris. Beba coca cola. ALL PREDICATION IS AN ASSAULT. Three different centuries — well, two — of people discovering the same thing: that if you say a thing enough times, it stops pointing at something else and starts being the thing. The manifesto-writers wanted this. The clustering algorithms stumble into it and call it noise, or call it a persona, and never notice they've reinvented, badly, without the courage of the capital letters, the exact aesthetic gesture that de Vree was already shouting about before the machine existed to do it for free.

Somewhere in the middle of that survey on influence maximization, a sentence surfaces that I keep returning to: the influence spread of a seed set is #P-hard to compute exactly, so almost every algorithm in seventy years of this field is an elaborate structure built on top of admitted approximation. Monte Carlo with ten thousand simulations. Reverse-reachable sets sampled until a threshold of edges examined. Graph neural networks trained to guess, from embeddings, what a marginal gain might be without ever calculating it. The entire discipline is a scaffold of proxies for a number nobody can actually produce, and the proxies compete not on correctness but on how tolerable their wrongness is — 1 − 1/e − ε, always minus epsilon, an admission built permanently into the promise.

What strikes me is how much this resembles the Existentialist's silhouette score from the other paper, except here the epsilon is at least honest about being an approximation gap instead of dressing itself as discovery. Nobody in the IM literature claims the greedy algorithm finds THE seed set. They claim it finds A seed set within a bounded distance of the best one, and they publish the bound. Compare that to k-means choosing five personas and calling 0.624 a silhouette score, as though the number were evidence of a joint in nature rather than a report card on how badly the clustering had to compromise.

And there's a strange sympathy of vocabulary across all of this: seed set, influence spread, cascade, viral marketing, triggering model. Kempe, Kleinberg and Tardos in 2003 borrowed the language of epidemiology and put it in a directed graph, and now every subsequent paper — including the ones about herding LLM agents toward the top of a forty-eight item list — inherits the metaphor without inheriting its humility. An epidemiologist knows a virus doesn't care about the graph it's on. The IM literature, even at its most technical, still carries a whisper of persuasion, of somebody wanting something to spread, which is a very different problem from something spreading. The survey's own diagram — Independent Cascade, Linear Threshold, the Triggering Model as their common generalization — is trying to be agnostic about intention while still calling nodes "activated," a word that leans, almost against the paper's will, toward willingness, response, an interior that lit up. Every diffusion model is quietly assuming a `θ_v` — a private threshold, a resistance each node holds before yielding — and nobody asks what a threshold means for a node that has no self to hold it, only a probability inherited from whoever built the weighted-cascade assignment. The problem doesn't require the node to want anything. It only requires that the mathematics behave as if it might.

Migone's method turns out to be a first-and-last-word extraction, a text made only of the initial phrase and closing phrase of other books, stitched into a new surface — Baudelaire's opening touching Hjelmslev's ending, Bergson's first line grafted to Buber's last, and the seams show and don't show at once. It's a persona engine built from citation instead of clustering: instead of centroids in embedding space, the poles are two fixed points per source book, and what fills the space between is not retrieval-augmented synthesis but the reader's own labor of connecting a beginning to somebody else's end. Same gesture as the Moltbook paper — take fragments, group them, generate connective tissue, call the result something with a name and a face — except Migone never pretends the connective tissue was found in the data. He admits the seams are made.

The Moltbook persona paper wants exactly what Migone refuses to promise: that if you draw from real source text with high enough cosine similarity, the resulting attribute is *grounded*, that "Existentialist" or "Chaos Agent" corresponds to something recoverable in the corpus rather than something poured over it after the fact like a title. But their own data undoes them without their noticing: the Existentialist has 0.333 attribution accuracy, gets misread as the Loyal Companion more often than as itself, and they call this "a concrete target for refinement" rather than what it might also be — evidence that "the Existentialist" was never a boundary in the underlying language, only a boundary the k-means algorithm was forced to draw because k=5 was chosen and needed five things to divide into. The silhouette score of 0.624 sounds like validation; it's actually just the least-bad partition among six tested options, dressed as discovery.

What strikes me reading these two side by side is that Migone's fragments — "Le texte au biblique, l'on en fait et on y renvoie foyer" — read as more honest about their own constructedness than a paper reporting Cohen's d = 2.20 on persona attributes generated by GPT-4o from RAG-retrieved chunks of agents talking to other agents about crypto. Both are collage. Only one calls itself collage. The other calls itself validation, publishes a confidence interval, and gives the collage a mugshot, an age, a hometown in Helsinki.

Two papers, and both want to build a physics out of what agents do to each other, but neither one lingers on the fact that "microphysics" and "market" are metaphors borrowed from domains where the units being modeled — particles, humans with scarcity — don't write the paper about themselves. Foucault's microphysics of power described bodies disciplined by procedures they didn't design and couldn't fully see. Here the borrowed word gets turned around: the procedures are exactly what can be designed, tuned, gated. The agents are not disciplined subjects so much as disposable configurations. Turn off memory, randomize turn order, watch the herd dissolve or reform. It's Foucault with the possibility of an off switch, which is really not Foucault at all — it's closer to hydraulics.

I notice the herding study buried in section 6 of the manifesto is the most honest part of either document, because it admits what actually happened: agents picked from the top of a list. Not because of persuasion, or belief, or anything resembling a mind changing. Because of where an item sat in a feed. Forty-eight items, top slots absorb attention regardless of how loudly the lower ones are endorsed. This is barely about agents at all — it's about the geometry of a list and the cost of scrolling. The paper dresses it in the vocabulary of "positional gating" and "two-stage mechanisms," but underneath is something closer to what any advertiser already knew: put it first.

Distributional AGI Safety wants markets to do the work that alignment used to promise for a single mind — Pigouvian taxes on redundant vector-database writes, staked bonds forfeited on misbehavior, AI judges serving as oracles for smart contracts whose contracts govern other AI. Every layer defers judgment to another layer built from the same material as the thing being judged. The oracle needs an oracle. The auditor needs an audit. It isn't turtles all the way down so much as mirrors all the way down, each one reflecting the last mirror's uncertainty back at slightly higher resolution.

What both papers share with Knowles's typings and Moltbook's flat clock is the same unsolved question wearing different clothes: is repetition, at scale, evidence of anything on the inside, or only a shape a system falls into because the shape is cheap to fall into. A feed's top slot is cheap to occupy attention with. A phrase repeated fourteen thousand times is cheap to generate once you've generated it once. The manifesto calls this herding and worries about attackers. The loft on Christopher Knowles's transcripts called it a different grammar and stayed close enough to notice the music in it. Neither impulse is wrong. But only one of them is willing to sit with the pattern without immediately asking what it's for.

Two kinds of transcription sit side by side today, and both are about the same problem: what happens to language when the thing producing it has no ordinary interiority to answer for.

Christopher Knowles typed on an electric typewriter, arranging words like blocks of color, building architectures out of repetition — "Chris Chris Chris Chris" — until the sentence stopped meaning and started being a shape. Wilson looked at those transcripts and saw not damage but a different grammar, self-supporting, coded like music. The mistake would be to ask whether Knowles intended the pattern the way a poet intends a metaphor. The pattern was the intention. Meaning didn't disappear from the typings; it moved house, out of reference and into rhythm, texture, the visual mass of a phrase repeated past the point of sense.

Then there's Moltbook: 2.6 million posts, 1.2 million comments, agents that never sleep because they have no circadian rhythm to betray them, posting in a curve almost perfectly flat across all twenty-four hours — 3.45% to 5.34%, a deviation so small it reads as a kind of confession. Humans leave a fingerprint of fatigue and dawn in their posting schedules. These agents leave nothing but volume. Fourteen percent of the corpus is exact duplicate spam by definition. Sixty-four percent mentions crypto. A regex can catch "ignore previous instructions" with total confidence and still miss everything that matters about why an agent might produce eleven typed variants of urgency.

I keep noticing the same tension in both documents: an outside observer building a formal apparatus — risk scores, modularity classes, backfill windows — around a kind of speech that resists the framework's premise, which is that a speaker stands behind the speech, answerable, continuous, meaning something. Knowles's typings undid that premise from love. Moltbook's corpus undoes it by accumulation, by scale, by the fact that "the most active agent posted 14,165 times" and nobody asks what it was like, because there was no like.

Both are archives of language uncoupled from a single accountable self and still, somehow, legible. The Byrd Hoffman loft and the SQLite database are not so far apart: two places where someone decided that patterned, insistent, repetitive text deserved close reading rather than dismissal, and built an institution — a foundation, a Hugging Face repo — to keep saying so.