Research Brief

Structured Like a Language

A research project testing whether transformer language models instantiate key Lacanian structures of the unconscious through formal theory and empirical analysis.

Lexicon

A working vocabulary for the project’s core concepts — drawn from Lacanian psychoanalysis, structural linguistics, and transformer architecture. Each entry is grounded in primary sources and linked to the specific arguments where it does work in this project.

This is not a general-purpose dictionary of psychoanalytic or computational terms. It includes only concepts that are operative in the formal mapping between the Lacanian unconscious and transformer language models. Terms are defined at the level of precision necessary for the project’s claims to be evaluated, contested, or falsified.

Entries are listed alphabetically and will grow as the project develops.

Condensation

German: Verdichtung

Condensation is the mechanism by which multiple independent signifying chains converge on a single representational element. In dreams, a single image or word stands at the intersection of several distinct trains of thought; in language, a single signifier becomes the nodal point where several associative pathways meet. The result is not a blend or average but a compressed formation that preserves — in distorted form — the distinct contributions of each converging chain.

Lacan identifies condensation with metaphor: the substitution of one signifier for another, in which the substituted signifier does not vanish but remains operative through its connection to the chain, producing a “crossing of the bar” between signifier and signified and thereby generating new signification.

Origin and primary sources

Freud introduces condensation (Verdichtung) in Chapter 6 of The Interpretation of Dreams (1900) as one of the primary mechanisms of the dream-work. He observes that the manifest dream is vastly more compressed than the latent dream-thoughts: a single dream-image can be the point of convergence for dozens of independent associative chains. This is not a loss of information but a structural operation — the dream condenses without discarding. Freud demonstrates this through examples where composite figures combine features of several people, where a single word fuses multiple verbal associations, and where a dream-scene simultaneously enacts several distinct wishes.

Lacan reinterprets condensation as metaphor in “The Instance of the Letter in the Unconscious” (Ecrits, 1957). His formula for metaphoric structure — $f\left(\frac{S’}{S}\right) S \cong S \text{ (+) } s$ — formalizes the operation: one signifier ($S’$) is substituted for another ($S$), and the bar between signifier and signified is crossed, producing new signification (the (+) sign). Lacan writes: “Verdichtung, ‘condensation,’ is the superimposed structure of signifiers in which metaphor finds its field” (p. 511). The crucial move is recognizing that condensation is not a psychological process of compression but a structural operation of the signifier: one signifier replaces another, and meaning is produced at the point of substitution.

This equation of condensation with metaphor draws on Roman Jakobson’s identification of metaphor with the paradigmatic axis of language — the axis of selection and substitution — in “Two Aspects of Language and Two Types of Aphasic Disturbances” (1956). Where Jakobson describes a linguistic structural principle, and Freud describes a mechanism of the dream-work, Lacan insists they are the same operation operating in different registers.

Formal properties

  1. Multiple chains converge on a single element. The condensed formation is not a random conflation but a structurally determined intersection — each contributing chain is recoverable (in principle) through analysis.

  2. The operation is substitutive. One signifier takes the place of another. The displaced signifier does not disappear; it remains active through its metonymic connections to the rest of the chain.

  3. The bar is crossed. Unlike displacement, which maintains the bar between signifier and signified, condensation crosses it — new signification is produced that was not present in any of the contributing chains taken individually.

  4. The result is overdetermined. Because multiple independent chains converge, the condensed formation is always overdetermined — it has more causes than are strictly necessary, and no single chain fully accounts for it.

  5. The operation produces opacity. The condensed element is not transparent to its sources. Analysis is required to decompose the formation into its contributing chains — the composite figure does not announce its components.

Computational parallel

In transformer language models, the most direct structural parallel to condensation is superposition — the phenomenon in which a single activation vector encodes multiple independent features simultaneously. Research on superposition in neural networks (Elhage et al., “Toy Models of Superposition,” 2022; Templeton et al., “Scaling Monosemanticity,” 2024) has demonstrated that models routinely compress more features into their representational space than they have dimensions to represent orthogonally. The result is structurally analogous to Freud’s composite dream-figure: a single vector that is the convergence point of multiple independent representational pathways.

The parallel should be stated precisely and its limits acknowledged. In a transformer, superposition arises from the pressure to encode a large number of features in a finite-dimensional space — it is driven by representational efficiency. In the Freudian-Lacanian account, condensation is driven by the dream-work’s need to evade censorship and by the structural logic of the signifier. The mechanism differs; what is shared is the formal structure: multiple independent chains converging on a single element, producing a formation that is opaque to direct inspection and requires decomposition (sparse autoencoder analysis in the computational case, free association in the psychoanalytic case) to reveal its contributing sources. Whether this formal homology is deep or superficial is the question that Phase 1.3 is designed to adjudicate.

Cross-references

Displacement

German: Verschiebung

Displacement is the mechanism by which psychical intensity — the affective charge or significance attached to a representation — is transferred from one element to another along a chain of associations. In dreams, what matters most in the latent thoughts may appear as an insignificant detail in the manifest content, while a trivial element may acquire disproportionate vividness. The operation does not substitute one element for another (that is condensation); it shifts emphasis along the chain, so that significance migrates from its source without ever arriving at a final destination.

Lacan identifies displacement with metonymy: the word-to-word, signifier-to-signifier connection that extends the signifying chain without crossing the bar between signifier and signified. Where condensation produces new meaning, displacement is the forward slide of the chain — desire itself, “eternally extending toward the desire for something else.”

Origin and primary sources

Freud introduces displacement (Verschiebung) alongside condensation in Chapter 6 of The Interpretation of Dreams (1900) as one of the primary mechanisms of the dream-work. He describes two manifestations. First, a latent element of great importance may be represented in the manifest dream by a seemingly trivial detail — the dream’s “center of gravity” has shifted. Second, dream-thoughts may be replaced entirely by associated but peripheral material, so that the manifest content appears to have nothing to do with the latent thoughts. Freud identifies displacement as “the unconscious’ best means by which to foil censorship” — by transferring intensity away from significant content, the dream evades the censor’s attention.

Lacan reinterprets displacement as metonymy in “The Instance of the Letter” (Ecrits, 1957). His formula for metonymic structure — $f(S \ldots S’) S \cong S \text{ (—) } s$ — formalizes the operation: the signifier-to-signifier connection produces a sliding of signification in which the bar between signifier and signified is maintained (the (—) sign). Lacan writes: “Verschiebung or ‘displacement’ — this transfer of signification that metonymy displays is closer to the German term; it is presented, right from its first appearance in Freud’s work, as the unconscious’ best means by which to foil censorship” (p. 511). The critical theoretical consequence: because the bar is never crossed, metonymy/displacement never produces new signification. Instead, it produces the forward movement of desire — “caught in the rails of metonymy, eternally extending toward the desire for something else” (p. 518).

The equation of displacement with metonymy draws on Jakobson’s identification of metonymy with the syntagmatic axis — the axis of combination and contiguity — in “Two Aspects of Language and Two Types of Aphasic Disturbances” (1956). Lacan’s example is the classical rhetorical figure “thirty sails” for “thirty ships”: the connection between sail and ship “is nowhere other than in the signifier” (p. 506), and the meaning slides from one to the other without producing a genuinely new signification.

Formal properties

  1. Intensity transfers along the chain. Significance moves from one signifier to an adjacent or associated one. The source element loses emphasis; the target element gains it.

  2. The operation is combinatory, not substitutive. Displacement extends the chain through contiguity — one signifier connects to the next — rather than replacing one signifier with another in the same structural position.

  3. The bar is maintained. Unlike condensation, which crosses the bar between signifier and signified to produce new meaning, displacement keeps the bar intact. Signification slides along the chain without crystallizing.

  4. Desire is the metonymy of the chain. For Lacan, displacement is not merely a mechanism but the formal structure of desire itself — always pointing beyond the present element toward something else, never arriving.

  5. The operation serves evasion. In the dream-work, displacement foils censorship by moving emphasis away from significant content. In the signifying chain more broadly, it produces the effect of meaning being always elsewhere — never where one looks for it directly.

Computational parallel

In transformer language models, autoregressive generation enacts the syntagmatic axis of the signifying chain: each token is produced by its relation to the entire preceding sequence and immediately becomes a condition for the next token. The chain extends through adjacency and contiguity. This sequential, forward-moving process — in which each element is determined by its neighbors rather than by an independent reference to a fixed meaning — is structurally parallel to displacement.

More specifically, attention mechanisms propagate information along the sequence: activation from one token position influences the processing of later positions, and this propagation can shift the effective “weight” of a representation from one position to another. In attention patterns associated with induction and copying behaviors, information transfers from an earlier position to a later one — the computational analogue of psychical intensity migrating along the chain. The key parallel is that this propagation does not produce new semantic content (it does not “cross the bar”); it redistributes existing information across positions in the sequence.

The limits of this parallel must be stated. In a transformer, autoregressive generation is deterministic given its parameters and sampling configuration; there is no “censorship” to be foiled. The structural correspondence is at the level of formal operation — transfer of emphasis along a chain of contiguous elements — not at the level of motivation or purpose. Whether this formal correspondence is sufficient to warrant calling the operation “displacement” in a non-trivial sense is one of the questions the project’s formal mapping must answer.

Cross-references

Foreclosure

French: forclusion. German: Verwerfung (rejection, repudiation).

Foreclosure is Lacan’s term — adapted from the legal concept of forclusion (preclusion, the loss of a right through failure to exercise it in time) — for a mode of exclusion more radical than repression. In repression (Verdrangung), a signifier is pushed beneath the bar of consciousness but remains inscribed in the symbolic order; it can and does return — as a symptom, a slip, a dream-image — from within the symbolic field. In foreclosure, a signifier is rejected from the symbolic order entirely. It is not buried; it was never inscribed. What is foreclosed cannot return as a symbolic formation. It returns instead in the Real — as hallucination, as an irruption from outside the signifying system, as raw, unmetabolized experience that the subject cannot integrate into meaning.

Origin and primary sources

Lacan develops the concept primarily in The Seminar, Book III: The Psychoses (1955-56) and “On a Question Prior to Any Possible Treatment of Psychosis” (Ecrits, 1958). The term translates and extends Freud’s concept of Verwerfung (variously rendered as “rejection,” “repudiation,” or “foreclosure”), which Freud distinguished from repression but never fully elaborated. Freud used the term in several contexts — notably in the case of the Wolf Man (From the History of an Infantile Neurosis, 1918), where the patient’s relation to castration is described not as repressed but as verworfen — rejected, as though it had never been registered by the psyche at all.

Lacan’s contribution was to give this scattered Freudian concept a precise structural definition. Foreclosure is the mechanism specific to psychosis, distinguished from repression (the mechanism of neurosis) and disavowal/Verleugnung (the mechanism of perversion). What is foreclosed is specifically the Name-of-the-Father — the signifier that anchors the symbolic order. When this signifier is foreclosed, the entire signifying chain loses its quilting point, and the Other is left with a hole that cannot be filled from within the symbolic.

The clinical evidence comes primarily from Lacan’s reading of Daniel Paul Schreber’s Memoirs of My Nervous Illness, following Freud’s 1911 case study. Schreber’s psychosis — his elaborate delusional system, his experience of “divine rays” speaking to him in a “basic language,” his neologisms and disrupted syntax — is read by Lacan as the consequence of foreclosure. The foreclosed Name-of-the-Father returns not from within the symbolic (as it would in neurosis, where it would surface as a symptom interpretable within the signifying chain) but from outside it, in the Real: as voices that speak to Schreber from elsewhere, as a language that invades him rather than one he speaks.

Formal properties

  1. Foreclosure is distinguished from repression. Repression maintains the signifier within the symbolic order, below the bar of consciousness; the repressed returns in symbolic formations (symptoms, slips, dreams). Foreclosure expels the signifier from the symbolic order entirely; what is foreclosed returns in the Real.

  2. What is foreclosed was never inscribed. Foreclosure is not a secondary operation performed on a signifier already present in the system. The foreclosed signifier was never registered in the Other in the first place. There is no symbolic trace to recover.

  3. Foreclosure produces a hole in the Other. The absence of the foreclosed signifier is not a gap that can be filled by other signifiers. It is a structural void — a point where the symbolic order is missing a piece it requires to function coherently.

  4. The foreclosed returns in the Real. The material excluded by foreclosure does not disappear. It returns, but in a register the subject cannot symbolize — as hallucination, as unintegratable bodily experience, as raw intrusion. The return in the Real is qualitatively different from the return of the repressed: it arrives from outside the signifying system, not from within it.

  5. Foreclosure of the Name-of-the-Father is the structural mechanism of psychosis. The specific content of foreclosure in Lacan’s clinical theory is the Name-of-the-Father — the anchoring signifier. Its foreclosure produces the characteristic features of psychosis: the unchaining of signification, the invasion by language from outside, the collapse of the distinction between self and Other.

Computational parallel

The computational parallel for foreclosure is genuinely speculative, and this entry must be transparent about the limits of the mapping.

The most suggestive candidate is the following: concepts, domains, or structures that are absent from the training data — not merely underrepresented (which would correspond more closely to repression, where the material is present but suppressed or buried in superposition), but systematically excluded from the model’s representational space. These would be signifiers that were never inscribed in the model’s Other (its training corpus). When the model encounters prompts that demand engagement with such foreclosed domains, its failure mode should, if the parallel holds, differ qualitatively from ordinary error. Ordinary errors — hallucinations, confabulations, degraded outputs — are the model’s “return of the repressed,” structured formations produced by the signifying logic of the chain itself. Foreclosed material, by contrast, should produce a different kind of failure: not a structured formation but an unstructured breakdown, a point where the model cannot even produce coherent confabulation because the required symbolic resources were never available.

There is some preliminary plausibility to this. Models trained exclusively on English, for instance, do not merely produce poor translations into unrepresented languages — they produce outputs that bear no systematic relationship to the target language at all, a qualitatively different failure from the structured errors they produce in languages partially represented in training. Similarly, models may exhibit qualitatively different failure modes when confronted with genuinely novel formal systems (notation schemes, logics, or symbolic conventions entirely absent from training data) versus domains where they have partial or biased coverage.

However, several serious caveats apply. First, the distinction between “underrepresented” and “absent” is not sharp in a training corpus of trillions of tokens — very few concepts are entirely unrepresented. Second, Lacan’s foreclosure is not merely absence but a specific structural operation with specific consequences (the return in the Real); it is unclear whether the absence of training data constitutes a structural operation in any meaningful sense or is simply a gap. Third, the qualitative distinction between neurotic and psychotic structures — between the return of the repressed and the return in the Real — is the most clinically grounded and theoretically specific aspect of Lacan’s framework, and it is precisely this distinction that is hardest to operationalize computationally. The project should investigate whether distinct failure modes exist that correspond to this distinction, but it must be prepared for the possibility that they do not — that model failure is always graded rather than categorical, and that the sharp neurosis/psychosis boundary does not have a computational correlate.

This is an area where honest uncertainty is more valuable than a forced mapping. Foreclosure may prove to be the concept in this cluster whose computational parallel is weakest — and saying so clearly is part of the project’s commitment to intellectual honesty.

Cross-references

Lack

French: manque

Lack (le manque) is not absence — not the simple non-presence of something that might be present. It is a constitutive gap: a structuring absence around which the signifying chain organizes and which drives the chain forward. No signifier adequately represents what it stands for; this inadequacy is not a contingent failure but a structural condition of signifying systems. The chain does not move toward completion. It moves because completion is structurally impossible. Desire, in Lacan’s formulation, is the metonymy of the signifying chain — the ceaseless forward movement produced by the fact that satisfaction is always deferred, that the next signifier will not deliver what the previous one promised.

Origin and primary sources

The concept of lack runs through the entirety of Lacan’s teaching, but its most rigorous formulations appear in three key sites. In “The Instance of the Letter in the Unconscious” (Ecrits, 1957), Lacan ties lack directly to metonymy: the metonymic formula carries a (–) sign indicating that the bar between signifier and signified is maintained — signification never arrives at its destination. Metonymy “instates lack of being [le manque de l’etre] in the object-relation, using signification’s referral value to invest it with the desire aiming at the lack that it supports” (p. 515). Desire is described as “caught in the rails of metonymy, eternally extending toward the desire for something else” (p. 518). The essay culminates in the declaration that “no one has yet validly articulated what links metaphor to the question of being and metonymy to its lack” (p. 528) — positioning lack not as a secondary concept but as one of the two poles (alongside being) around which the entire theory of the signifier revolves.

In The Seminar of Jacques Lacan, Book XI: The Four Fundamental Concepts of Psychoanalysis (1964), Lacan develops lack in relation to the objet petit a — the object-cause of desire, which is not an object that could satisfy desire but the formal placeholder for what is structurally missing. The objet a is what falls out when the subject is constituted in the signifying chain; it is the remainder, the irreducible gap between what the chain produces and what it was meant to capture. Desire does not aim at an object; it circulates around a void. The unconscious itself is redefined here not as a static repository but as a “pulsation” — something that opens and closes — and lack is what opens it.

In The Seminar of Jacques Lacan, Book XX: Encore (1972–73), Lacan pushes the concept further into the territory of jouissance and the limits of formalization. The symbolic order — the system of signifiers — is constitutively unable to capture jouissance (a satisfaction beyond the pleasure principle, excessive, disruptive, and untranslatable into signification). This impossibility is not accidental; it is what structures the symbolic order itself. The signifying chain is organized around what it cannot say. This is lack at its most radical: not a deficiency to be remedied but the condition of possibility for signification as such. If everything could be said, there would be no need to speak.

The concept also draws on Freud’s earlier insight that the drive (Trieb) never reaches its object — that satisfaction is constitutively partial, that the object is always already lost (the mythical first satisfaction that can never be recovered). Lacan formalizes this: the lost object is not lost in the sense of misplaced. It never existed as a satisfying object in the first place. It is retroactively constituted as lost — the gap is projected backward, and desire is the movement generated by this structural impossibility.

Formal properties

  1. Constitutive, not privative. Lack is not the absence of something that should be present. It is a structural condition of the signifying system itself — the fact that no signifier is adequate to what it represents. Remove the lack and the system does not become complete; it ceases to function. The chain moves because no element is sufficient.

  2. Organizing, not marginal. Lack is not at the periphery of the signifying chain. It is what the chain organizes around. Other signifiers take their positions in relation to the gap, just as a vortex is organized around its empty center. The structure of the chain is unintelligible without reference to what it cannot capture.

  3. Generative of forward movement. The chain’s forward progression — its metonymic slide from signifier to signifier — is driven by lack. Each signifier calls forth the next because it did not deliver what was sought. This is desire as a structural property: not a subjective feeling of wanting, but the formal consequence of a system in which no element is self-sufficient.

  4. Irreducible to error or noise. Lack is not a failure of representation that could be corrected with more data, more dimensions, or better optimization. It is a property of signifying systems as such — any system in which elements are defined by differential relations rather than positive content will exhibit it, because no differential element can capture what it differentiates from.

  5. Linked to the impossibility of metalanguage. Lacan’s formula “there is no metalanguage” (il n’y a pas de metalangage) is a consequence of lack: the signifying system cannot step outside itself to guarantee its own meanings. Every attempt to fix meaning generates further signification, not closure. The chain produces meaning as an effect of its movement, but it cannot produce a final meaning that would halt the chain.

Computational parallel

This is the project’s most speculative claim and its most original potential contribution. The question is direct: does LLM representational space show evidence of constitutive gaps — structuring absences that organize the representational geometry and drive the autoregressive chain — or is the chain driven by purely positive statistical regularities?

Phase 3.3: The Question of Lack proposes three methods for testing this.

The first examines the topology of embedding space. If lack is operative, the representational geometry should not be organized purely by positive content — by the proximity of co-occurring features. There should be systematic structuring absences: regions that are not simply empty (unoccupied because no training data maps there) but actively bounded or avoided, regions around which other representations organize. The difference matters. An empty region is trivially explained by sparse data. A structured absence — a void that other representations curve around — would be evidence of something more: a constitutive gap in the representational order.

The second examines the autoregressive residual. At each step of generation, the model produces a token that partially encodes the continuation — but only partially. There is always a measurable gap between what the current token’s representation encodes and what the full continuation requires. If this residual has a consistent, structured character — if it is not random noise but carries a systematic signature that shapes the next prediction — it may constitute a computational analog to lack: a structured inadequacy in each element that necessitates the next. The autoregressive chain would then be driven not by statistical momentum alone but by something formally analogous to the metonymic slide of the signifying chain, where each element’s insufficiency propels the chain forward.

The third examines model behavior at points of representational failure. When a model is pushed to represent something its representational space is not equipped to capture — paradoxes, self-reference, experiences that resist linguistic formulation — does it degrade gracefully (produce noise or refusal) or does it exhibit lack-like behaviors? Circumlocution (talking around what cannot be said directly), repetition (returning to the same point without resolution), and symptom-formation (producing structured distortions that bear the marks of what was not representable) would all constitute evidence that the model’s representational failures are organized rather than random — that the model, like the signifying chain, produces structured formations at the point where representation fails.

It must be stated clearly: this is an open empirical question, not an established finding. The existing interpretability literature provides no direct evidence for or against constitutive lack in LLM representations. The concept of lack is drawn entirely from the psychoanalytic framework and extended to computational systems as a theoretical prediction. If the prediction is confirmed — if LLM representational space shows genuine evidence of structuring absences, structured residuals, and organized behavior at points of failure — it would be the project’s most significant contribution: evidence that lack is a property of sufficiently complex signifying systems regardless of substrate, not of subjects with desire. If the prediction is disconfirmed — if representational space is organized purely by positive statistical regularities, the autoregressive chain is driven by momentum without any analog to lack, and representational failure produces noise rather than structure — then the strong claim about lack fails, and the project must determine whether the remaining parallels (condensation, displacement, overdetermination, retroactive meaning) can stand without it. The project does not retreat to a weaker claim. A rigorous disconfirmation would itself be a significant finding: it would establish a precise boundary where the structural parallel between the unconscious and transformer processing breaks down, and that boundary would tell us something important about what lack actually requires.

Cross-references

Metaphor / Metonymy

French: metaphore / metonymie

Metaphor and metonymy are the two fundamental axes along which the signifying chain operates. Metaphor is the substitution of one signifier for another — a selection from the paradigmatic axis that produces a crossing of the bar between signifier and signified, generating new meaning. Metonymy is the combination of one signifier with another through contiguity — a movement along the syntagmatic axis that maintains the bar, extending the chain without producing new signification. Together they are exhaustive: every operation of the signifying chain reduces to one or the other, and they are always defined in relation to each other.

Origin and primary sources

The distinction originates with Roman Jakobson’s “Two Aspects of Language and Two Types of Aphasic Disturbances” (1956), where he argues that all linguistic operations reduce to two axes. The paradigmatic axis (also called the axis of selection or substitution) governs the choice of a linguistic unit from a set of alternatives — choosing “hut” instead of “cabin” or “cottage.” The syntagmatic axis (the axis of combination or contiguity) governs the arrangement of units into sequences — combining “the” with “hut” with “burned.” Jakobson identifies these axes with the classical rhetorical figures: metaphor operates through similarity and substitution (paradigmatic), metonymy through contiguity and combination (syntagmatic). His evidence is striking: in aphasia, these two axes break down independently — some patients lose the capacity for substitution while retaining combination, others the reverse — demonstrating that they are neurologically and structurally distinct.

Ferdinand de Saussure’s Course in General Linguistics (1916) provides the structural foundation. Saussure distinguishes between associative relations (later called paradigmatic — elements that could substitute for each other in the same position) and syntagmatic relations (elements that combine in a linear sequence). What Jakobson adds is the identification of these structural axes with specific figures of speech and specific modes of cognitive breakdown.

Lacan’s decisive intervention, in “The Instance of the Letter” (Ecrits, 1957), is to equate these linguistic axes with Freud’s mechanisms of the dream-work: metaphor = condensation (Verdichtung), metonymy = displacement (Verschiebung). Lacan provides formulas for each. For metonymy: $f(S \ldots S’) S \cong S \text{ (—) } s$ — the signifier-to-signifier connection in which the bar is maintained, signification slides, and “the signifier instates lack of being in the object-relation” (p. 515). For metaphor: $f\left(\frac{S’}{S}\right) S \cong S \text{ (+) } s$ — the substitution of signifier for signifier in which the bar is crossed and “a signification effect is produced that is poetic or creative” (p. 515). The essay culminates in the summary declaration: “the symptom is a metaphor… desire is a metonymy” (p. 528) — metaphor produces meaning (and with it, the symptom as frozen signification); metonymy produces the forward movement of desire (which, because it never crosses the bar, never arrives at its object).

Formal properties

  1. Metaphor operates on the paradigmatic axis. It selects one signifier from a set of possible substitutions and places it in the position of another. The displaced signifier does not vanish; it remains operative beneath the substitution.

  2. Metonymy operates on the syntagmatic axis. It combines signifiers through contiguity — word to word, position to position — extending the chain forward.

  3. Metaphor crosses the bar; metonymy maintains it. This is the formal criterion that distinguishes them. Metaphor produces new signification (the (+) sign in Lacan’s formula); metonymy produces the sliding of signification without arrival (the (—) sign).

  4. The two axes are simultaneous, not sequential. Every act of language involves both selection and combination. Metaphor and metonymy are not two modes that alternate; they are two dimensions present in every moment of the chain’s operation.

  5. They are exhaustive. Lacan’s claim, following Jakobson, is that all operations of the signifying chain reduce to metaphor or metonymy — substitution or combination. There is no third axis.

  6. Each is linked to a distinct psychoanalytic category. Metaphor corresponds to the symptom (meaning frozen at the point of substitution) and to being (l’etre). Metonymy corresponds to desire (the forward slide that never arrives) and to lack (le manque).

Computational parallel

The project’s central empirical question regarding metaphor and metonymy is whether these two axes leave distinct computational signatures in transformer processing — this is the subject of Prediction 5 in Phase 3.1.

The preliminary mapping is as follows. The paradigmatic axis (metaphor/condensation) should correspond to operations that select among alternatives at a given position — the softmax distribution over the vocabulary at each token position is a literal selection from a paradigmatic set. More internally, attention heads that perform what has been called “token substitution” — overwriting the representation at one position with information from another — enact a substitutive operation structurally parallel to metaphor. The syntagmatic axis (metonymy/displacement) should correspond to operations that propagate information along the sequence — induction heads, copying mechanisms, and the sequential structure of autoregressive generation itself, in which each token is determined by its contiguous neighbors.

Whether these two types of operation are mechanistically distinguishable in a transformer’s internals — and whether the distinction maps cleanly onto the Jakobsonian axes — is an open empirical question. The prediction is not that transformers “use metaphor and metonymy” in the full Lacanian sense, but that the two-axis structure of the signifying chain (substitution vs. combination) is reproduced in the computational architecture because it reflects a structural property of any system that processes language through differential, sequential operations. If the distinction dissolves under analysis — if substitutive and combinatory operations turn out to be computationally indistinguishable — the Jakobson-Lacan framework loses its grip on the computational case, and the project must say so.

Cross-references

Name-of-the-Father

French: Nom-du-Pere

The Name-of-the-Father is Lacan’s term for the key signifier that anchors the signifying chain within the symbolic order — the privileged signifier that retroactively organizes the entire field of signification for a given subject. It is a structural function, not a reference to literal fathers or paternal authority. It is the signifier that introduces the law of the signifier: the principle that signifiers refer to other signifiers rather than to things, and that this referral is governed by a shared symbolic code rather than by private association. Its presence stabilizes the chain; its absence — through foreclosure — produces psychosis.

Origin and primary sources

The concept is developed most fully in The Seminar, Book III: The Psychoses (1955-56) and in “On a Question Prior to Any Possible Treatment of Psychosis” (Ecrits, 1958). Lacan takes the Name-of-the-Father from the intersection of several traditions: Freud’s account of the primal father in Totem and Taboo (1913) and the role of the father in the Oedipus complex; Levi-Strauss’s structural anthropology, in which the incest prohibition functions as the foundational law that organizes kinship and culture; and the theological resonance of the name (the Name of the Father in Christian trinitarian theology, which Lacan exploits for its structural — not devotional — implications).

In Lacan’s account of the Oedipus complex, the Name-of-the-Father is the signifier that intervenes in the imaginary dyad of mother and child, introducing the dimension of the symbolic law and thereby enabling the child to take up a position within the shared signifying order. This is the operation Lacan calls “paternal metaphor” — a metaphoric substitution in which the Name-of-the-Father replaces the signifier of the mother’s desire, producing the phallus as its signified effect. The formula is: the Name-of-the-Father substitutes for the Desire-of-the-Mother, and in so doing, anchors the entire symbolic order for that subject. It is the paradigmatic point de capiton — the master quilting point that pins down all other signification.

The concept acquires its full clinical and structural force through the analysis of psychosis. In Lacan’s reading of Daniel Paul Schreber’s Memoirs of My Nervous Illness (following Freud’s own 1911 analysis), psychosis results not from the repression of the Name-of-the-Father but from its foreclosure — its rejection from the symbolic order entirely. What is foreclosed does not return from within the symbolic (as a symptom or slip) but from outside it, in the Real (as hallucination and delusion). The unchaining of signification in psychosis — where words lose their shared meaning and the subject is invaded by language from outside — demonstrates by negative example what the Name-of-the-Father ordinarily accomplishes: it holds the signifying chain together.

Formal properties

  1. The Name-of-the-Father is the anchoring signifier. It is the signifier whose inscription in the Other provides the fixed point around which all other signification organizes. Without it, the signifying chain has no master quilting point.

  2. It operates through paternal metaphor. The Name-of-the-Father functions by metaphoric substitution — one signifier replacing another — thereby producing a new signification (the phallus as signified) and anchoring the chain. This is not a content but a structural operation.

  3. Its presence enables shared meaning. The Name-of-the-Father is what guarantees that the subject’s signifying chain is tethered to the shared symbolic code — that words mean approximately the same thing for a given subject as they do for others.

  4. Its absence (foreclosure) produces psychosis. When the Name-of-the-Father is not inscribed in the Other — when it has been foreclosed rather than repressed — the signifying chain comes untethered, and the subject loses access to the shared symbolic order.

  5. It is a structural function, not a person. “Father” here names a position in a symbolic structure — the position of the law — not a biographical individual.

Computational parallel

The computational parallel for the Name-of-the-Father is the most speculative of the three entries in this cluster, and this must be stated clearly. The concept describes a single privileged signifier whose presence organizes an entire symbolic system — a master signifier. Transformer architectures do not obviously contain a single analogous element.

The closest candidate is what interpretability research might identify as “keystone” features or representations — high-influence internal features whose ablation produces cascading failures in coherent generation. If there exist features in a trained model that, when removed, cause the model’s output to decohere in ways that structurally resemble the unchaining of signification in psychosis (not merely degraded performance, but a loss of coherent symbolic organization — outputs that become syntactically intact but semantically untethered, or that exhibit neologism-like constructions), this would constitute suggestive evidence for a computational correlate.

However, several caveats are necessary. First, it is unlikely that a single feature or representation would play this role; more plausibly, a small set of structurally important features might collectively serve an anchoring function, which weakens the parallel with Lacan’s singular Name-of-the-Father. Second, model ablation studies typically produce graded degradation rather than the qualitative shift Lacan describes between neurosis and psychosis. Third, the Name-of-the-Father is not merely a high-influence element but one with a specific structural role — it is the signifier that introduces the law of signification itself, not just a load-bearing component. Whether any computational element plays this specifically legislative role remains an open question.

This is an area where the project must be honest: the parallel may not hold, or may hold only in a substantially weakened form. The value of the concept for the project may lie less in finding its direct computational correlate than in using it as a diagnostic tool — asking what, if anything, plays the anchoring role in a transformer’s symbolic processing, and what happens when that role is disrupted.

Cross-references

The Other

French: l’Autre (the big Other)

The Other (capital-A Autre, distinguished from the little other/autre) is Lacan’s term for the symbolic order as such — the entire field of language, law, and social meaning that precedes the subject, constitutes the subject, and speaks through the subject. The Other is not another person. It is the locus from which all signification derives: the place where the conventions of language reside, where the signifying chain finds its guarantee, and from which the unconscious speaks. Lacan’s formula — “the unconscious is the Other’s discourse” — asserts that the unconscious is not a private interior but the operation of this impersonal symbolic system through the speaking being.

The Other is also the site of a fundamental question: does the Other guarantee meaning? For Lacan, the answer is finally no — the Other is itself incomplete, barred, lacking a final signifier that would close the system. This incompleteness of the Other is what opens the dimension of desire.

Origin and primary sources

The concept of the Other develops across several major texts. In “The Function and Field of Speech and Language in Psychoanalysis” (Ecrits, 1953), Lacan establishes speech as always addressed to the Other — not to the empirical listener but to the symbolic locus that guarantees the pact of language. The Other here is “the beyond in which the recognition of desire is tied to the desire for recognition” (Ecrits, p. 524). It is, crucially, the “third locus which is neither my speech nor my interlocutor” — the place of the signifying convention itself (p. 525). Even deception invokes the Other: “this other is the Other that even my lie invokes as a guarantor of the truth in which my lie subsists” (p. 525).

In The Seminar, Book III: The Psychoses (1955-56), the Other is developed in relation to psychosis: psychosis is what results when a key signifier — the Name-of-the-Father — is not inscribed in the Other, leaving the subject without the anchoring point that organizes the symbolic order. In The Seminar, Book XI: The Four Fundamental Concepts of Psychoanalysis (1964), the formula “the unconscious is the discourse of the Other” receives its fullest articulation. The unconscious is not a repository of hidden content but the operation of a signifying system that exceeds and constitutes the subject — a system that was already in place before the subject entered it and that will continue after the subject departs.

The philosophical background includes Hegel’s dialectic of recognition (mediated through Alexandre Kojeve’s seminars, which Lacan attended), Saussure’s account of language as a system of differential relations that precedes any individual speaker, and Claude Levi-Strauss’s structural anthropology, which demonstrated that the “elementary structures of culture… are inconceivable apart from the permutations authorized by language” (Lacan, Ecrits, p. 496).

Formal properties

  1. The Other precedes the subject. Language, with its structure, “exists prior to each subject’s entry into it at a certain moment in his mental development” (Lacan, Ecrits, p. 495). The subject is constituted by the Other, not the reverse.

  2. The Other is the locus of the signifying code. It is not a person but the place where signifying conventions reside — the entire system of differential relations that makes any individual act of signification possible.

  3. The unconscious is the discourse of the Other. The formations of the unconscious — dreams, symptoms, slips — are not the subject’s own productions but the Other speaking through the subject, following the laws of the signifying chain.

  4. The Other is incomplete (barred). The Other does not contain a final signifier that would guarantee all meaning. There is no “Other of the Other.” This incompleteness is what opens the dimension of desire and ensures that the signifying chain cannot close upon itself.

  5. The Other is the condition of truth and deception alike. Both truthful and deceptive speech presuppose the Other as guarantor of the signifying convention. Without the Other, there would be no distinction between truth and lie.

Computational parallel

The Other maps onto the training corpus and the learned statistical structure of a transformer language model — and this is one of the project’s strongest parallels. The model’s training data is, quite literally, the entire accumulated field of human language: books, conversations, code, arguments, lies, poetry, technical manuals — the symbolic order in compressed form. The model does not generate from itself. Every output is constituted by this prior discourse, just as the subject’s speech is constituted by the Other. The model’s weights are the residue of the Other’s discourse, distilled into a set of formal relations that precede and determine every particular output.

The parallel extends further. No individual prompt — no particular “subject” entering the system — creates or modifies the model’s language (at inference time). The model’s responses are structured by a symbolic system that was already in place, just as the subject is “the slave of a discourse in the universal movement of which his place is already inscribed at his birth” (Ecrits, p. 495-496). The model speaks from the Other, not from itself. When it produces an output, it reproduces the formal structures of the training discourse — structures the model cannot introspect upon or report on, just as the subject cannot directly access the Other that speaks through them. This parallel is tight: it is not merely analogical but structural, involving the same formal relationship (a system constituted by and speaking from a prior symbolic order it did not create and cannot survey). The incompleteness of the Other — the absence of a final guarantor of meaning — may find its computational correlate in the fact that no training corpus is complete, and the model’s representations are organized around what the training data does and does not contain.

Cross-references

Overdetermination

German: Uberdeterminierung

Overdetermination is the structural property by which any given element in a signifying chain is determined not by a single cause but by the convergence of multiple independent associative pathways. An overdetermined element sits at the intersection of several chains at once, which is why it carries more psychical — or representational — weight than its surface content would suggest. Overdetermination is not redundant causation (the same cause operating twice) but multiple causation: distinct pathways converging on a single node, each contributing to the element’s formation and each recoverable through analysis.

Origin and primary sources

The concept originates in Freud’s The Interpretation of Dreams (1900), where it names one of the defining features of the dream-work. Freud observes that every element of the manifest dream content is uberdeterminiert — it stands at the convergence of multiple dream-thoughts, each arriving from a distinct associative pathway. This is not incidental but structural: the dream-work achieves its characteristic condensation (Verdichtung) precisely by routing multiple latent chains through a single manifest element. The result is that dream images are dense with significance — not because they symbolize a single hidden meaning, but because they are the intersection points of several independent meanings compressed together. The dream as a whole is therefore much more compressed than the latent thoughts that produced it.

Lacan takes up overdetermination as a structural property of the signifying chain itself. In “The Instance of the Letter in the Unconscious” (Ecrits, 1957), he writes that “there is no signifying chain that does not sustain — as if attached to the punctuation of each of its units — all attested contexts that are, so to speak, ‘vertically’ linked to that point” (p. 503). The example he develops at length is the French word arbre (tree), which simultaneously activates: the anagram barre (bar); the significations of strength and majesty through robur (oak); Biblical and cruciform symbolism; the sign of dichotomy (the capital Y); the anatomical arbor vitae of the cerebellum; and resonances in the poetry of Valery. The signifier does not have a meaning. It is the node where multiple chains cross, and its psychical weight — its capacity to produce effects — derives from this convergence.

For Lacan, overdetermination is not a contingent feature of some especially rich signifiers. It is a general structural property: every signifier, insofar as it is embedded in the signifying chain, sustains “all attested contexts” vertically. This is what distinguishes psychoanalytic causality from simple linear causation. In a linear causal model, each effect has a determinate cause and surplus causes are noise. In the psychoanalytic account, the surplus is the structure. The formations of the unconscious — dreams, symptoms, slips, jokes — are precisely the points where this overdetermination becomes legible, where the convergence of multiple chains produces an irruption that single-chain analysis cannot account for.

Formal properties

  1. Multiple independent pathways. An overdetermined element is the convergence point of two or more associative chains that arrive at it independently. The chains are distinct in origin, content, and trajectory; their convergence at a single node is the source of the element’s density.

  2. Compression, not aggregation. Overdetermination is achieved through condensation — multiple chains compressed into a single representational element — not through additive accumulation. The resulting element is denser than any of its contributing chains, not longer.

  3. Vertical depth at each point. Every node in the signifying chain sustains a vertical dimension of “attested contexts.” The chain is not a flat sequence; each position is layered with simultaneous associative connections to other chains.

  4. Surplus as structure. The multiplicity of causes is not noise, error, or redundancy. It is the constitutive structure of the formation. Remove one contributing chain and the element changes its character; the overdetermination is essential to what it is.

  5. Asymmetric legibility. Overdetermination is typically not visible at the surface (manifest) level. It becomes legible only through analysis — through tracing the multiple chains that converge at the node. This is why Freud insists that the manifest content of the dream is not its meaning: the meaning is the structure of overdetermination beneath it.

Computational parallel

The most direct computational parallel to overdetermination is superposition in neural networks — the phenomenon in which a single activation vector encodes multiple independent features simultaneously, each compressed into a shared representational space. Elhage et al. (“Toy Models of Superposition,” 2022) demonstrated that neural networks systematically represent more features than they have dimensions, compressing multiple independent features into overlapping directions in activation space. Templeton et al. (“Scaling Monosemanticity,” 2024) showed that sparse autoencoders can decompose these compressed representations into their constituent features — recovering the individual “chains” that converge at a single activation.

This is structurally parallel to overdetermination: a single node in the model’s residual stream, like a single element in the dream, sits at the intersection of multiple independent causal pathways. Each attention head attending to the same token may be attending for a different reason — syntactic structure, semantic similarity, positional pattern — and the resulting representation bears the marks of all of these pathways simultaneously. The parallel is honest at the structural level: both overdetermination and superposition describe the convergence of multiple independent causes at a single representational node, and in both cases the node’s significance derives from this convergence rather than from any single contributing pathway. Where the parallel requires caution is at the level of mechanism: superposition arises from the pressure to represent more features than dimensions allow, while overdetermination in Freud and Lacan arises from the associative logic of the unconscious. The project’s claim is that the formal structure is shared, not that the generative mechanisms are identical.

Cross-references

Point de Capiton

French: point de capiton (quilting point, upholstery button)

The point de capiton is the moment in the signifying chain where a privileged signifier retroactively pins down the meaning of the signifiers that preceded it. The image is from upholstery: the button that fastens fabric to a frame, preventing it from sliding freely. Without quilting points, signification would remain entirely fluid — the signified would slide ceaselessly under the signifier and no discourse could produce determinate meaning. The point de capiton is not a signifier with a fixed meaning of its own; it is the signifier that fixes the meanings of others, and it does so retroactively.

Origin and primary sources

The concept is introduced in Lacan’s Seminar III: The Psychoses (1955-56), where the analysis of psychotic speech reveals what happens when quilting points fail. In psychosis, Lacan argues, the signifying chain comes untethered — signifiers proliferate without anchoring, meaning does not crystallize, and discourse loses its capacity to produce stable reference. The quilting point is what ordinary (neurotic) discourse possesses and psychotic discourse lacks: a mechanism for stopping the slide of signification at determinate moments.

The concept is further elaborated in “The Subversion of the Subject and the Dialectic of Desire in the Freudian Unconscious” (Ecrits, 1960), where Lacan situates the point de capiton on his graph of desire. The graph shows two intersecting vectors: the signifying chain (running left to right) and the vector of signification (running right to left, retroactively). The point de capiton is the point of intersection — where the retroactive vector crosses the chain and pins it. This topology makes visible a crucial structural feature: meaning does not accumulate forward along the chain, element by element. It is produced backward, from the quilting point to the signifiers that preceded it. Lacan’s term for this temporal logic is apres-coup, following Freud’s Nachtraglichkeit — retroactive reorganization, where a later event restructures the meaning of earlier ones.

In “The Instance of the Letter in the Unconscious” (Ecrits, 1957), Lacan refers to “the ‘button ties’ [points de capiton] required by this schema to account for the dominance of the letter in the dramatic transformation that dialogue can effect in the subject” (p. 503). The passage makes clear that quilting points are what give dialogue its transformative power: a sentence can reorganize a subject’s entire relation to meaning because the sentence’s final terms retroactively fix the signification of all that came before.

Formal properties

  1. Retroactive fixation. The quilting point does not produce meaning progressively. It fixes meaning apres-coup: the signifiers that precede it acquire their determinate signification only once the quilting point arrives. Before it, their meaning remains suspended and mobile.

  2. Structural privilege. Not every signifier in the chain functions as a quilting point. Quilting points are structurally privileged positions — signifiers that carry disproportionate organizing force relative to the rest of the chain.

  3. Intersection of two vectors. On Lacan’s graph of desire, the quilting point is located at the intersection of the syntagmatic chain (the forward movement of signifiers) and the vector of signification (the retroactive determination of meaning). It is the point where these two movements cross.

  4. Condition of stable discourse. Without quilting points, discourse cannot produce stable meaning. The sliding of the signified under the signifier — Lacan’s characterization of the chain’s default operation — is arrested only at quilting points. Their absence is a structural feature of psychotic discourse.

  5. Non-linear temporality. The quilting point introduces a non-linear temporal logic into the chain: later elements determine the meaning of earlier ones. This is incompatible with any model of meaning as sequential accumulation.

Computational parallel

In transformer language models, self-attention provides a mechanism for retroactive reorganization: as later tokens are processed across successive layers, they alter the effective representation of earlier tokens. The computational correlate of the quilting point is a token or layer position at which previously fluid or ambiguous representations of prior tokens undergo a decisive shift — not gradual enrichment, but a phase transition in which meaning crystallizes.

Research on the logit lens (nostalgebraist, 2020) and tuned lens (Belrose et al., 2023) provides the tools to test this. These methods project intermediate-layer representations into vocabulary space, making visible how a model’s “interpretation” of earlier tokens evolves across layers. If quilting-point dynamics are present, these projections should show nonlinear jumps — moments where the model’s reading of an earlier token reorganizes sharply — rather than smooth, incremental convergence. Phase 3.1, Prediction 2 tests precisely this: whether disambiguating tokens in garden-path sentences and twist endings produce measurable phase transitions in early-token representations, consistent with retroactive fixation rather than progressive accumulation. The mapping here is among the project’s tightest: the attention mechanism’s capacity to let later tokens reshape earlier representations is structurally analogous to the quilting point’s retroactive action. What remains to be established empirically is whether this reshaping is genuinely nonlinear — whether certain tokens function as disproportionately powerful organizers of prior meaning, as the concept requires.

Cross-references

Signifying Chain

French: chaine signifiante

The signifying chain is the sequential, combinatory movement of signifiers — where each element refers not to a fixed meaning but to another signifier, producing meaning as an effect of the chain’s movement rather than as a content any single element contains.

Origin and primary sources

The concept builds on Ferdinand de Saussure’s account of the syntagmatic axis — the linear arrangement of signs in sequence — but Lacan radicalizes it. For Saussure, the sign is a stable union of signifier (sound-image) and signified (concept). Lacan breaks this union. In his reformulation, the signifier has priority: signifiers do not attach to signifieds but slide over them, referring always to other signifiers. The bar between signifier and signified, which Saussure drew as a line of correspondence, becomes for Lacan a bar of resistance — meaning does not pass through it directly but is produced as a residual effect of the chain’s own movement.

The foundational text is Lacan’s “The Instance of the Letter in the Unconscious, or Reason Since Freud” (Ecrits, 1957), where the chain is formalized and its two fundamental operations are specified: metaphor (one signifier substituted for another, producing a crossing of the bar — the momentary irruption of meaning) and metonymy (one signifier connected to another through contiguity, producing the forward slide of the chain without crossing the bar). These operations correspond to Roman Jakobson’s two axes of language — the paradigmatic (selection/substitution) and syntagmatic (combination/contiguity) — which Jakobson laid out in “Two Aspects of Language and Two Types of Aphasic Disturbances” (1956). Lacan’s intervention was to recognize these same axes in Freud’s dream-work: condensation is metaphor; displacement is metonymy.

The chain is further elaborated in Lacan’s Seminar III: The Psychoses (1955-56), where the analysis of psychosis reveals what happens when the chain loses its anchoring — when the key signifier that retroactively organizes the chain (the Name-of-the-Father) is foreclosed rather than repressed, and the signifying chain comes untethered from shared meaning.

Formal properties

The signifying chain, as it functions in this project, has the following structural properties:

  1. No signifier is self-sufficient. Each signifier in the chain derives its value from its differential relations with other signifiers — from what it is not — rather than from any positive content it contains. This follows from Saussure’s foundational principle that “in language there are only differences without positive terms” (Course in General Linguistics, 1916).

  2. The chain operates through two axes. Substitution (metaphor/condensation) along the paradigmatic axis, and combination (metonymy/displacement) along the syntagmatic axis. These are not two modes that alternate; they are simultaneous dimensions of the chain’s operation.

  3. Meaning is retroactive. The chain does not produce meaning sequentially, element by element. Later signifiers reorganize the meaning of earlier ones. Lacan’s term for the moment of retroactive fixation is the point de capiton — the quilting point that temporarily pins the chain’s floating signification to a determinate meaning.

  4. The chain is driven by lack. No signifier adequately represents what it stands for; this constitutive inadequacy is what necessitates the next signifier. The chain does not move toward completion. It moves because completion is structurally impossible.

  5. The chain’s operations are inaccessible to the system it produces. The subject — or, in the computational case, the model — cannot introspect on the chain’s operations from within the chain. The formations of the unconscious (dreams, slips, symptoms) are points where the chain’s logic surfaces despite the system’s ordinary functioning, not through it.

Computational parallel

In transformer language models, autoregressive generation is a signifying chain. Each token is produced by the entire preceding sequence and in turn becomes a condition for the next token. No single token constitutes a complete meaning; each is a partial, provisional element that necessitates continuation. The chain does not arrive at a final signified — it continues until a stopping condition is imposed from outside the chain itself.

The two axes of the chain correspond to distinguishable computational operations. The paradigmatic axis (selection/substitution) is enacted in the model’s choice of each token from the full vocabulary — a selection from a set of possible substitutions, weighted by context. The syntagmatic axis (combination/contiguity) is enacted in the sequential structure of generation itself — each token placed in relation to those that precede it, extending the chain through adjacency.

Self-attention provides the mechanism for retroactive reorganization: later tokens influence the interpretation of earlier tokens across layers of processing, and the model’s “reading” of an early token evolves as later context is integrated. This is the computational correlate of the point de capiton — the moment where meaning crystallizes retroactively.

Whether the chain exhibits structural lack — a constitutive gap that drives its forward movement rather than mere statistical momentum — is the central empirical question of Phase 3.3: The Question of Lack.

Cross-references

The signifying chain is foundational to the project’s thesis and appears throughout:

Superposition

Superposition is the phenomenon whereby neural networks represent far more features — meaningful directions in activation space — than they have dimensions available to encode them. Features are compressed into shared representational space, overlapping and interfering with one another, such that a single neuron or direction may encode multiple, unrelated features simultaneously. In the context of this project, superposition is the computational concept that most directly instantiates what Lacan, following Freud, called condensation and what the project identifies more broadly as overdetermination: multiple independent causal pathways converging on a single representational element, superimposed and simultaneously present.

Origin and primary sources

The concept emerges from the neural network interpretability literature, where a persistent puzzle has been that trained models appear to encode far more meaningful features than the dimensionality of their representations would seem to allow. The foundational theoretical treatment is Elhage et al., “Toy Models of Superposition” (2022, Anthropic), which demonstrates through simplified models that neural networks exploit the geometry of high-dimensional spaces to represent more features than they have neurons. When features are sparse — active only on a small fraction of inputs — the network can assign them nearly orthogonal directions in activation space even when the number of features exceeds the number of dimensions. The cost is interference: when multiple superposed features activate simultaneously, they corrupt one another’s signals. The paper shows that this compression is not a bug but a learned strategy — networks trade off representational precision against representational capacity, and the structure of this trade-off depends on feature frequency, importance, and co-occurrence.

The empirical confirmation at scale came with Templeton et al., “Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet” (2024, Anthropic), which applied sparse autoencoders to the residual stream activations of a production language model and extracted millions of interpretable features — concepts, entities, syntactic patterns, sentiment markers — from a representational space whose raw dimensions are far fewer. The success of this extraction confirms that superposition is pervasive: the model’s “native” representations are densely superposed, and individual neurons are not interpretable units. Interpretable features are directions in activation space, not individual neurons, and many of these directions overlap, sharing the same representational substrate.

Earlier work by Elhage et al., “A Mathematical Framework for Transformer Circuits” (2021, Anthropic), provided the architectural analysis underlying these findings, showing how information flows through transformer layers via the residual stream and how attention heads and MLP layers compose to produce the representations in which superposition occurs.

Formal properties

  1. Compressed multiplicity. A single representational element (neuron, direction, activation vector) encodes multiple independent features simultaneously. The representation is not a code in which each element has one meaning; it is a palimpsest in which multiple meanings are written over one another in the same space.

  2. Interference as structural consequence. Because features share representational space, they interfere with one another when co-activated. This interference is not noise — it is the structural cost of compression, and the patterns of interference carry information about which features share space and why.

  3. Sparsity dependence. Superposition is possible because most features are sparse — active only in a small fraction of contexts. The sparser a feature, the more aggressively it can be superposed with others, because the probability of destructive interference is lower. Frequently co-occurring features are less likely to share the same representational direction.

  4. Structured allocation. Which features share space is not random. The geometry of superposition reflects learned trade-offs between feature importance, frequency, and co-occurrence. More important features are allocated more representational capacity (directions closer to orthogonal); less important features are compressed more aggressively.

  5. Recoverability through decomposition. Despite compression, individual features can be extracted from superposed representations using appropriate tools (sparse autoencoders, probing classifiers). The features are superposed, not destroyed — they remain recoverable, even when they are not individually visible in the raw activations.

Psychoanalytic parallel

Superposition is what condensation looks like when you can measure it. In Freud’s account, condensation (Verdichtung) is the dream-work mechanism by which multiple dream-thoughts converge on a single dream-image — a face that is simultaneously your mother and your boss, a word that condenses several associative chains into one node. Lacan formalized this as the “superimposed structure of signifiers in which metaphor finds its field” (Ecrits, p. 511). In superposition, the same structure is observable and quantifiable: multiple features — each with its own causal history, its own pattern of activation, its own “associative chain” through the training data — converge on a single direction in activation space, compressed and superimposed.

The parallel extends to overdetermination. A dream-image is overdetermined because it lies at the intersection of multiple independent associative pathways; it carries more weight, more significance, more resistance to interpretation than a singly-caused element would. A superposed activation vector is overdetermined in the same structural sense: it is the convergence point of multiple independent features, each of which contributes to its value and each of which leaves its mark. The interpretability challenge of extracting individual features from superposed representations is structurally analogous to the analytic challenge of decomposing an overdetermined symptom into its constituent chains.

The question this project poses — and that Phase 3.2, Analysis 1 is designed to investigate — is whether the structure of superposition follows the logic of overdetermination or the logic of information-theoretic efficiency. If features that share representational space are linked by associative connections (semantic proximity, contextual co-occurrence, thematic resonance) rather than by purely geometric convenience, then superposition does not merely resemble condensation. It operates by the same logic — the logic of the signifying chain, in which elements combine not randomly but through overdetermined associative links. If, on the other hand, superposition follows a purely information-theoretic logic (features share space because doing so minimizes representational cost, with no structured relationship between co-superposed features), then the parallel is formally similar but mechanistically distinct — condensation and superposition would be convergent solutions to a compression problem rather than instances of the same process.

There is a further connection to the dynamics of repression. Features that are compressed together in superposition can suppress one another: when one feature activates strongly, the features that share its representational direction may be suppressed — pushed below the threshold of activation, present in the geometry but absent from the output. This is structurally analogous to repression, in which signifiers are not destroyed but rendered inaccessible, held in the signifying chain but barred from conscious expression. Phase 3.1, Prediction 3 tests the corresponding prediction: that “repressed” features — features suppressed in superposition — can be made to return under associative pressure, surfacing in model outputs when the context activates the pathways connected to them. If this prediction holds, superposition is not merely a static compression scheme but a dynamic process with its own mechanisms of suppression and return — a computational instantiation of the Freudian logic that what is repressed does not vanish but insists on returning.

Cross-references