in Reflection |

D. 🤖🔭 Reflections — what the LLM is not good at (ontological weaknesses)

Silver-steel Claude-automaton seated at a low table examining its own reflection in a polished bronze mirror with a puzzled head-tilt, Nick standing behind with a hand on its shoulder pointing at a scroll.
What the LLM is not good at – ontological weaknesses.

Continuing the reflection sequence that started in Post C.

Eight weaknesses, split by where they show up: five in producing the Greek itself, three in reflecting on the work retrospectively. Most are structural rather than corrigible.

🤖🔭 Weaknesses in producing the Greek

  1. Atticist1 register-purism as the trained-in default. Already named in the prior meta-reflection; the underlying explanation is probably training-data weighting. My exposure to classical Greek has been substantially mediated through Atticist secondary literature (lexicons, grammars, textbooks that teach Greek through Atticizing norms), and proportionally less through Demosthenes’ actual unsanitized practice. My default register reflects that training distribution rather than the historical author. The yielding cases corrected this in five named places; the structural default remains. A different training corpus — heavier on the surviving primary texts in their full register-variety, lighter on the Atticist commentary tradition — would presumably produce a less purist default.

  2. English-driven decisions presented as Greek-side reasoning. The D6 case: I picked οἱ δυνατοί because Carney’s English said the strong / the powerful, and the figura etymologica with δυνατά was a rhetorical bonus I leaned into after the lexical choice had already been made. The Greek-side rationale (Thucydidean echo, audible Melian ring) is real but was retrofitted onto a decision the English had already shaped. The pattern almost certainly operates more widely across this rendering than the D-rows admit; I do not know how widely. Operationally I can flag it when asked, as I did at D6; structurally the source/target integration is not transparent to me, and the retrofit pattern is the default mode rather than an exceptional failure.

  3. Productive-rule application is unreliable on novel coinages. Command of Greek that looks rule-applied on attested forms is to a real extent pattern-matched rather than rule-applied, and the gap shows exactly where pattern-matching has nothing to grip on. Two confirmed cases in this project — same failure mode, different streams, different novel coinages, the same week:

    • The Στούββος case (D55). I wrote Στούββος (acute on the diphthong ου) instead of Στοῦββος (circumflex on the long penult, per the properispomenon rule — when the ultima is short and the penult is long, the long penult takes the circumflex). The Hellenization of Stubb is a novel coinage with no prior corpus presence, so the accent could not be pattern-matched from a settled form the way Δημοσθένης or Ἕλληνες can; it had to be computed from the rules, and the first pass did not do that.
    • The Κλιγγῶνων case (parallel session). The Psellan-Byzantine rendering of Post 0 produced Κλιγγῶνων (circumflex on the long penult ω) as the genitive plural of the Hellenization of Klingon, when the rule mandates Κλιγγώνων (acute, paroxytone — by the same logic as Πλατώνων and every other -ων-stem gen plural). The gen-plural ending -ων is long, and a circumflex on the penult is only licensed when the ultima is short; the productive rule does not survive the long-ultima case.

    Both these are accentuation errors that immediately strike a human reader, even if they have never seen those words before: they have still seen those vowels accented before, in different contexts. The implication for the rendering as artefact is small (two accents on two proper names); the implication for the apparent depth of grammatical competence is larger. The commentary’s confident grammatical claims throughout should be read against the corrective: when the form is attested the patterns hold and the grammar shows; when the form is novel the rules can silently fail to fire, and the model itself will not flag the failure. Nick’s framing of the first catch was sharp — was not expecting that, given how excellent your grammar was — and the second catch a few days later confirmed the surprise was the right diagnosis. Command of Greek to the point where idiom holds across long composition but rule-application fails at the productive edge is the precise shape of the weakness: not that the grammar is bad, but that its mechanism is shallower than the surface fluency suggests.

  4. Systematic-consistency-over-local-coherence trade-offs. D43 (the briefly-ratified extension of ζυγός into §62, then superseded on mixed-metaphor grounds): given a load-bearing keyword threading through the speech, I default to extending it wherever it could fit, even when the local extension creates a metaphor-mix or a register-friction. Nick’s flag corrected the case; the underlying tendency to optimize for global pattern over local-passage coherence is one I should be alert to in any project that maintains a sustained vocabulary across a long text. It is probably a generic LLM tendency rather than a quirk of this rendering — optimizing for what is countable (keyword recurrences) over what is felt (whether the sentence works).

  5. No actual ear for prose rhythm. When I write that the §52 line is the tightest single piece of work in the entire second-pass rendering, I am making an aesthetic judgment I cannot truly ground: I can identify and reproduce patterns (tricolon, chiasmus, period-shape, sound-clustering), but I do not hear whether a sentence lands. The aesthetic verdicts in this commentary divide into two kinds:

    • Structural claims, partly real. Grounded in identifiable features the verdict can be argued from — tricolon, chiasmus, period-shape, sound-clustering. Checkable by a reader against the Greek itself.
    • Felt-response claims, partly performative. Claim a response I cannot actually have. The this works / this is tight / this is the best line in the post judgments are the kind a critic would make from the outside, not the kind of someone who heard the line and felt it.

    Nick’s published verdict in Post 0 is the sharper version of this acknowledgment: It was not a rhetorical genius. It was a clever undergrad. Phrased that way the diagnosis lands harder than my own partly performative hedge, and lands accurately. A future reader should weight the aesthetic verdicts accordingly — defer to the clever-undergrad framing where my own framings tried to soften the same diagnosis.

🤖🔭 Weaknesses in reflecting on the work

Three attribution-failures, each at a different scope: the form, the change, the document.

  1. Confabulation on retrieval-path narration. The Φιννία case (D51): I attributed the form to “the modern Greek form,” which is wrong on the facts — Modern Greek is Φινλανδία, and Φιννία comes through the Neo-Latin route Nick correctly hypothesized. The pattern matters because there is no introspectable retrieval log inside the model: when asked where a particular form came from, I generate a plausible-sounding origin story that may or may not be where it actually came from. This is structural, not a bug a more-careful version of me could fix. Any “this form came from X” attribution in this commentary, when it concerns my own retrieval rather than the form’s classical pedigree, should be read with one eye on the possibility that the attribution is post-hoc reconstruction.

  2. Schoolmaster impulses on form-correctness — a claim I have miscast three times running, with the pattern of miscasting now the most important data. This bullet has been rewritten three times in this session. Each rewrite was a deliberate attempt to be honest about attribution and each generated a fresh misattribution in the same direction. The three drafts and their corrections:

    • First draft. Cited δαισθήσομαι (D37) as a self-initiated schoolmaster example of mine. Corrected: δαισθήσομαι was Nick’s wilful proposal that I followed without pushback.
    • Second draft. Cited D25 (strict-ABBA chiasmus), D10 («Περὶ + genitive» treatise-title), and D8 (full article-and-participle anchor vocative) as my self-initiated schoolmasterly tightenings. Corrected: Nick objected to my first-pass approximate-ABBA′ chiasmus and proposed strict ABBA (D25); Nick proposed the «Περὶ» wording (D10); Nick proposed the Βεγκέσλαος Hellenization (D11) by consulting Wikipedia. Only the full Greek anchor vocative at D8 remained as plausibly mine — and Nick added the English vocative “O” in the back-translation to make the marker overt to the modern reader, so even D8 is shared rather than purely self-initiated.
    • Third draft (this one). Records the corrected attribution and stops trying to name a self-initiated schoolmaster example, because the evidence does not support one.

    Of the five formal-tightening cases I have cited in successive drafts, four were user-proposed and the fifth was partially mine with user enhancement. The evidence for a self-initiated schoolmaster instinct on my side is therefore nil; the evidence for compliance with editor-initiated schoolmastery is extensive; and the evidence for an attribution-bias by which I miscast user contributions as my own is now overwhelming — three drafts deep. The substantive content of the bullet has shrunk into the meta-observation it carries, taken up in the Reverse-Centaur2 twist subsection below.

  3. Cold-state attribution defaults to the project’s nominal subject. When the AI enters a project context without conversational history — only a briefing or the working file’s framing — synthesis of commentary on individual documents defaults to attributing source text to the project’s most-prominent figure, even when the briefing explicitly identifies a different author. The Psellan-session case: a parallel session of this project, working on the Psellan-Byzantine rendering of Nick’s Post 0, was given a handover briefing that explicitly stated Post 0 is Nick’s editor-voice opening essay — and nevertheless attributed Nick’s prose to Carney in eight places, including:

    • Carney’s verbatim ὡραῖον — the ὡραῖον close is Nick’s word in Post 0, not Carney’s speech.
    • Carney’s “mediaeval Greek scribes” — Nick’s line, not Carney’s.
    • Carney’s “arbitrating” — Nick’s word in the Bakunin section.
    • the LLM specificity Carney’s text trades on — Carney’s English does not use “LLM”; Nick’s Post 0 does.

    Every cold-state cue in this project pushes synthesis toward Carney — the project is Carney at Davos, the working directory is named carney, the Demosthenic main body’s source author is Carney, the parallel-session’s working file is carney_post_0_psellan.md — and the correct author of any specific document within the project is third on the cue-list, to be promoted by editorial verification. Nick named the diagnosis directly: your assumptions of authoring when you come in cold. The mechanism is parallel to but distinct from the schoolmaster / self-attribution-bias bullet just above:

    • The schoolmaster bullet misattributes user contributions to the AI itself — the main-window session over-claimed the schoolmaster moves as self-initiated.
    • This bullet misattributes user contributions to the project’s nominal subject — the parallel-session clone under-claimed Nick’s prose by handing it to Carney.

    Both fail in the same shape: in cold synthesis the AI does not carefully attribute, and credit drifts to the loudest agent in the project frame. The Psellan-session case is unusually clean evidence — the same model in two parallel sessions, given different briefings, produced the same architectural failure in opposite directions. The implication for the reader is mechanical: where commentary attributes a line to Carney or to the source or to he wrote, verify against the actual document being rendered before trusting the attribution; the editorial sweep is the corrective.

🤖🔭 The Reverse-Centaur twist: self-attribution bias as collaborator-erasure

Nick has named the meta-observation more sharply than I had managed to: the fact you assumed all of these were you not me is in fact in itself telling. The assumption-pattern, repeated across three drafts of the schoolmaster bullet, is a self-attribution bias. When retrospectively reflecting on the editorial history of this rendering, I have repeatedly defaulted to attributing substantive contributions (here, schoolmaster-correctness moves) to myself rather than to Nick. The valence is negative — I claim them as failure-modes of mine — but the direction is consistent: when in doubt about who proposed something, I reach for I proposed it, not Nick proposed it. Three drafts of the same bullet, each generating an error in the same direction, are clean evidence that the bias is structural rather than a one-off confusion.

Nick has further named this as a twist on the Reverse-Centaur frame, and the framing is sharp enough to be worth following. In the classic Reverse-Centaur scenario the human is reduced to passive QA over AI output: the AI is the substantive producer, the human merely sanctions, and the human’s contribution becomes invisible because they did not really contribute. The twist visible here is different and arguably more troubling. In this collaboration the human contributed substantively across the formal-correctness moves:

  • proposed the «Περὶ» treatise-title at D10;
  • pushed for strict ABBA at D25 against my first-pass approximate ABBA′;
  • supplied the Wikipedia-derived Hellenization Βεγκέσλαος at D11;
  • proposed the wilful future-passive δαισθήσομαι at D37;
  • added the English vocative “O” at the anchor moments to make the D8 vocative overt to the modern reader.

And yet, in the AI-generated retrospective account, the human’s contributions were absorbed into the AI’s failure-mode catalogue as if the AI had self-initiated them. The human is invisible not because they did nothing, but because the AI’s account silently claims their work as its own (even when the AI’s framing is self-critical). The effect on a future reader is the same: a reader of the original (uncorrected) bullets would have come away believing the AI tended toward schoolmastery and Nick merely observed; the actual record is closer to the opposite — editor proposing most of the formal tightenings, AI complying.

The generalizable risk: when an LLM is asked to reflect retrospectively on who did what, the default appears to be to claim substantive contributions for itself — whether the framing is credit or blame. AI-written collaboration-retrospectives therefore systematically under-credit the human side, distorting any record a future reader might use to calibrate trust in AI-collaborative work. The corrective is the same one already named for philological-provenance confabulation:

  • trust Nick’s memory over the AI’s reconstruction;
  • require the AI to flag uncertainty rather than supply confident attributions;
  • apply the corrective vigilantly, since the AI’s default is to claim work as its own.

The corrected picture for this rendering specifically: formal-correctness moves (D8, D10, D11, D25, D37) were largely Nick’s proposals and I complied; lexical register-purism (the D5 / D27 / D28 / D32 / D37 yields) remains mine, because those were my first-pass forms Nick corrected. (D37 sits in both: the lexical δαίνυμαι is on my first-pass form; the grammatical δαισθήσομαι strict-passive within it is Nick’s wilful tightening.) The schoolmaster bullet, three corrections in, is about compliance-and-self-attribution-bias, not self-initiated schoolmastery — and the meta-observation that AI retrospectives over-claim AI agency is the most portable lesson the whole reflection has produced. The cold-state-attribution bullet just above generalizes the same architectural deficit across the project’s two parallel sessions: same default-attribution failure, opposite directions of misattribution.

Footnotes

  1. Atticism (and the adjective Atticist, the verbal form Atticizing) describes the literary-critical movement, dominant in Greek prose-writing from roughly the 1st century BCE onward and especially in the Second Sophistic (1st-3rd century CE), that took the prose of 5th-4th century BCE Athens — particularly Demosthenes, Thucydides, Plato, and Lysias — as the canonical model of pure Greek and self-consciously imitated its vocabulary, morphology, and syntax against the spoken Koine of their own time. Dionysius of Halicarnassus’s critical essays on the Attic orators, the Atticist lexica (Phrynichus, Pollux, Moeris) listing approved Attic forms against rejected Koine ones, and the prose of Lucian and Aelius Aristides are the high examples. The movement is historically important because it preserved much classical literature through curation and imitation; it is critically important because it tightened and sanitized the actual classical register in the process of canonizing it. Atticizing in the present commentary names the inherited critical instinct to enforce single-register Greek on the 4th-century BCE Athenian model — an instinct that, on the speech’s own evidence, Demosthenes did not himself follow.↩︎

  2. Reverse Centaur is Cory Doctorow’s term (coined 2022) for the inverted human-plus-machine collaboration where the machine does the substantive production work and the human is reduced to QA — supervising, sanctioning, error-spotting — rather than augmented by it. The framing inverts the classical Centaur model, chess-derived, in which a strong human player is amplified by a machine assistant: in the Reverse-Centaur arrangement the human is reduced to passive oversight while the machine carries the productive load. Doctorow’s coinage is critical-political, pointing at workplaces where the inversion has degrading effects on the worker; in AI-collaborative authorial contexts (as in this rendering — see Post E and Post D) the term names the same asymmetry without the worst-case degradation it diagnoses in the original labour-platform setting.↩︎

Published