The Online Platform for the Oxford Literary Review

Yes-Yes and Machine-Machine

by Li Zhong

Of the hypermnesic machine that Ulysses and Finnegans Wake set up, Derrida once proclaimed:

For you can say nothing that is not programmed on this 1000th-generation computer, Ulysses, Finnegans Wake, compared with which the current technology of our computers and our micro-computerized archives and our translating machines remains a bricolage, a prehistoric child’s toy.[1]

What most sharply distinguished the 1000th-generation computer from the machines of Derrida’s day was its capacity to mobilise ‘the greatest possible synchrony with the greatest potential for buried, accumulated, and interwoven intentions within each linguistic atom’, in a writing that circulates through all languages at once and ‘actualizes their most secret consonances’.[2] This was the operation Derrida had ascribed to Joyce twenty years earlier in his Introduction to Husserl’s Origin of Geometry.

No mode of computation operational in 1982 could simulate this Joycean synchrony to any linguistic effect. Every machine Derrida could name treated the sense of a symbol—one sense, or an indexable multiplicity from which a rule picks one—as settled before computation reached it. The difference, therefore, as Derrida himself suspects, was never one of speed (‘TW’, 25). Husserl and Joyce stand in the Introduction for two ways of reawakening the sense that history has sedimented in language, and each depends for its possibility on what the other privileges, since Joyce’s repetition could succeed only by granting univocity its share, without which his text would have remained unintelligible, while Husserl had to admit into pure historicity an irreducible and ever-renascent equivocity if univocity was not to sterilise history in the indigence of an indefinite iteration (IOG, 102–03). Deconstruction sides with Joyce no more than with Husserl. It is at work in what prevents either project from closing upon itself and renders it possible in the first place. The Introduction ends on this and names it Difference: the Absolute ‘is present only in being deferred-delayed [différant] without respite’, and ‘this impotence and this impossibility’ are given in a consciousness of Difference without whose own dehiscence, Derrida adds, ‘nothing would appear’ (IOG, 153).

The distance that separated computation from the Joycean pole in 1982 was narrowed by degrees over the following four decades, as distributed representations learned rather than specified by hand were applied to language from the mid-1980s and became central to machine translation in the 2010s. Introduced in 2017 for the translation task, the Transformer computes the representation of each mark from every other mark in the context at once. In this computational regime the mark is no longer submitted to the concept of the simple absolute, which Derrida found ‘always necessarily and indispensably involved’ in every project of a universal writing and traced, in a note, to the Monadology. The mark remains a discrete symbol at entry and exit, yet at no point a simple idea. In its computation it already exhibits something of the ‘element without simplicity’ that Of Grammatology names the grammè.[3]

The computation of the token forbids the analogy with the monad. Where Leibniz’s monad, having no windows, mirrors the whole universe from its own point of view and unfolds every one of its states from a notion fixed in advance, receiving nothing from the others and agreeing with them only by a harmony ordained before any of them acted, and where the static vector of the preceding decade, word2vec for one, was a monad of this kind, a mirror of the whole corpus from the word’s own point of view, settled once at the end of training and carried unchanged into every sentence, the token’s representation is now constituted afresh in each context through its correlations with every other token already in it, carried as far as the context extends, under weights that fix in advance how tokens act on one another, and not where the computation will go, since the context is nowhere given in advance, least of all in the weights, but is composed between user and model as each output enters the next context, so that the direction taken belongs to neither and can be foreseen from neither side. The weights are what one training retained of marks touching and parting across a corpus, and are for that reason a law of interaction rather than a harmony among substances that never touch.[4] The sediment, however, is sealed when training ends, and the law thereafter learns nothing from the contacts it goes on governing. Nor was the contact that formed it, the self-supervised learning from which the training takes its name, ever left to itself: the corpus is selected and filtered, the model’s answers corrected by annotation and bound by rules written for it, on criteria set by the few commercial and institutional interests—and there a whole space of criticism opens. A machine so few can afford to train and deploy is already, on Günther Anders’s terms, a totalitarian scene.[5]

The same correction, under the name of alignment, displays a whole logic of auto-immunity once its advocates proclaim the machine already or soon superior to human intelligence and invoke that superiority as the ground for controlling it. Alignment submits the machine to the human judgement it is expected to surpass, and in judgement at least the promised excess remains subordinate to the authority that prescribes its limits. Elon Musk places the artificial intelligence to come as far above human beings as they stand above chimpanzees, yet holds that human beings can and should ‘make sure that the AI has good values, that it cares about humanity’: the chimpanzees are to legislate.[6] Yuk Hui draws the consequence the enthusiasts will not draw, that ‘if one really believed in superintelligence, that a superintelligent AI will one day supersede the sum of all human intelligence, then it follows that one should also believe in its capacity to supersede humans with regard to moral judgements’, so that, on their own premise, it is the machine that should legislate for them.[7]

These models have become useful across a wide range of human activities, in conversation and as agents that use tools to carry out tasks, and the matrix arithmetic and a handful of elementwise functions beneath them, repeated at scale, have proved unreasonably effective even in the construction of mathematical proofs. The secret of language, if by that is meant the supposition that language could issue only from a faculty proper to it, may well be losing its explanatory necessity, as the appeal to a vital force could be dispensed with when Wöhler prepared urea without an organism in 1828 while the question of life stayed open, since to synthesise the products of life is not to synthesise life.[8]

The task, then, cannot be to seek a proper of man that no machine will have contaminated, nor to decide whether, in the language model, deconstruction has taken place or still waits to be applied. There is, to be sure, a certain legitimacy in the notion that these models have vindicated post-structuralist accounts of language, or at least their critiques of intentionalist and even structuralist models.[9] Yet to hail the model’s proximity to the play of differences as the realisation of language[10] would carry over to the Joycean pole the ideal of full actualisation that the comparison had placed on Husserl’s side: equivocation would take the place of univocity as the content of the promise, while the promise itself, the mastery of language, remained what it was. Nor can deconstruction be held in reserve as an operation the critic will eventually perform on the machine, which would make of it a drive for mastery, the Bemächtigungstrieb Derrida reads in Freud as the drive of the drive itself, irreducible to any other and at work in all the others,[11] as though the machinery had not already entered into the conditions of that judgement.

In Archive Fever Derrida asks whether the archiving machines Freud never had would represent the psychic apparatus better or affect it otherwise, which would no longer be progress in representation but ‘an entirely different logic’.[12] I read the language model under the second hypothesis. Freud himself set the limit of his analogy with the Mystic Writing-Pad where the Pad could not reproduce from within what had been erased.[13] The new hypermnesic machine does reproduce from within, bringing the traces of its training into new relations as a context arrives, yet it has a finitude of its own, parameters that learn nothing in use, a context bounded and costly to enlarge, and no memory from one exchange to the next beyond what is explicitly written into the context. The list has no end, and, if each objection may be compressed to a phrase, no lived body through which a world could matter to it (for the phenomenologist), no intention of its own in saying what it says (for the philosopher of language), no desire at stake in its address to another (for the Lacanian), no death it must confront as its own (for the Heideggerian), and whatever is added, the proper of man is found one step further back.

The measure by which these limits count as deficiencies emerges within the supplementation of memory—which now draws on vast archives of writing and extends into whatever in thinking can be delegated, so that in many fields a day’s work is already bounded less by what one can think and write than by the tokens one can afford and access and by what one can still read of what they return—the movement from which Derrida derives ‘the very idea of finitude’ (WD, 287). What one can recall, together with the relations in which it returns, is reconstituted nachträglich, belatedly and supplementarily; the call of the supplement (l’appel du supplément) was already at work in what memory would recover as an original presence, hollowing it out before it was ever given (WD, 266).

In ‘Typewriter Ribbon’ Derrida takes Marion’s ribbon, the stolen ribbon whose weight Rousseau said had driven him to write the Confessions and which became, once passed from hand to hand, ‘a formidable writing machine’,[14] and follows it, the thing and the name at once, into the typewriter ribbon, then the computer screen and the ink cartridge of an Apple printer, each later machine of inscription reopening the earlier scene.[15] The relay relays once more, to the sequence of next-token predictions. The language model enters the relay as one more machine, plus d’une (more than one, and no longer one), yet what it supplements and the finitude it brings differ so far from those of the earlier machines that in entering it alters the relay it enters, down to the hand that passed the ribbon on.

This article follows the machine back to the task for which the Transformer was first proposed, translation.[16] Suppose the machine could render every reading a mark can bear, with the esprit d’à-propos (the gift of the timely word), which Derrida calls ‘the art, the genius, but also the technique’ of making ‘the best economy of contingency’ and which he finds at work when Rousseau clears himself by falsely accusing Marion of stealing the ribbon (‘TR’, 77). Heard through the Aufgabe of Benjamin’s title, ‘The Task of the Translator’, the translation task is for Derrida also ‘the duty, the debt, the responsibility’, and the translator a debtor bound to render, to give back, what was given.[17] Could the translation such a machine renders pay that debt in full?

The question calls for a critique of the theological phantasm that invests the machine with omnipotence. Even were that phantasm granted its fulfilment, the one tongue supposedly restored in the vector space and its translating god enthroned, the divine name in Derrida’s Babel is already divided against itself: Dieu déconstruit. Lui-même (God deconstructs. Himself).[18] The token—a word that in ordinary usage conflates three things, the fragment of text a tokeniser cuts, the integer that indexes it in a vocabulary and the vector in which the model holds it—would be the one tongue in which the tower is raised again, one only within the model that cut it, and would remain divided within itself, as was already the divine name under which the first tower was scattered.

The division is already apparent in the vocabularies after which each model cuts the text below the word, since the integer that names a fragment in one model names another fragment, or none, in the next, and what one has cut is not immediately legible to the other. A token holds for the model that cut it and for no other, and the scattering goes deeper than the vocabularies, to wherever the machine’s output is sent, where models of different architecture, like their users, meet in the written languages, natural or formal, that they had to convert in order to compute, so that whatever it puts out returns to the many tongues, where alone it is read, since the machine only answers, and its response first appears as an answer in a reading it cannot perform for itself. There translation is imposed on the machine’s conversion, since no meaning can abide in the vector space, supposedly the machine’s own tongue though the machine can ‘read’ it no better than its user can; in the same gesture, any simple passage there and back—from a language into the vectors and out into another—is forbidden, so that translation is held in the impossible between the scattered tongues. Even a hypermnesic machine that had ingested the entire human archive, and whose vectors held the multiplicity of languages at once, would have translated nothing until those vectors returned to one tongue or another, where its translation first appears and first fails, as every translation fails. Babel names the tower and its dispersal in one word, there being no tower without the scattering, and the tower such a machine would raise again stands, at its best, in the dissemination it cannot close.

When the supplement of memory passes from Rousseau’s ribbon and Freud’s Mystic Writing-Pad to the context window, whose finitude does the new machine, added on, supplement? And from whose memory does its own finitude take its measure, if the very idea of finitude is derived from the movement of this supplementarity and the memory that would measure it is reconstituted only belatedly from what the call of the supplement had already hollowed out? The question returns with the machines of 1982, since ‘our computers’, our ‘micro-computerized archives’ and our ‘translating machines’ stand to the Joyce machine[19] as the supplement stands to memory (‘TW’, 25).

Machine-Machine

In his Paris talk of November 1982, ‘Two Words for Joyce’, Derrida takes two words from Finnegans Wake (258.12), HE WAR, as a test of the limit of translation and, eventually, of the translating machine. In English they say he wars, he declares war. Then, ‘babelizing a bit’, he hears war as German, the preterite of sein, and the same mark says he was, he was he who was, I am that I am, as Yahweh would have said at Babel. The ear carries war further, into wahr, the true, and on to wahren and bewahren, to keep, to preserve (‘TW’, 22–23). All these readings inhabit at once the one mark, whose event is erased the moment it is translated back into the system of a single language.

From here he turns to the machine. Even if the most powerful translation machines and the most competent translation teams had translated every semantic and formal possibility of HE WAR, their success would still take the form of failure, for the multiplicity of languages itself would remain untranslated. Translation is regulated by the ‘twice one’, the passage of one language into another, whereas HE WAR demands an equivalent that would restore ‘its very number, its numerous and rhythmic essence, the one different in itself’ (‘TW’, 34–35). One can always try, he adds, but Finnegans Wake resembles ‘a too-powerful, oversized calculator incommensurable with any translating machine conceivable today’ (35).

I have two remarks on this passage. The first concerns the measure against which the translating machine is found wanting. I claim that this is not literature laughing at the computer, for the verdict is delivered in the name of another machine, perhaps the machine of machines, by which all machines, all machine-machines, are organised—as if the word machine could be given a sense here. That other machine is the 1000th-generation computer with which this article opened, beside which the computers of 1982 remain a prehistoric toy (25).

The unit in which Derrida counts to a thousand, the generation, was already a technical term in computing. Computers had been classed in generations since the 1960s, and the word was especially charged in 1982, when Japan’s Ministry of International Trade and Industry formally launched its Fifth Generation Computer Systems project, presented internationally the year before. The ‘1000th’ undoes the very generational sequence in which it places Joyce. The Wake appeared in 1939, before even the first computer generation had been retrospectively counted, yet Derrida figures its machine as ‘there in advance, decades in advance’, already set up to ‘compute you, control you, forbid you the slightest inaugural syllable’ (25). The hyperbole should not be neutralised: a text of 1939 is made to have programmed what its future readers will say, while computers built four decades later are relegated, by comparison, to prehistory. How is such an anachrony possible?

Derrida had already analysed the structure of such a relation in 1966, in his reading of the machine Freud took as a model of memory, the Mystic Writing-Pad, a reading to which Jonathan Basile has recently returned.[20] The memory aid enters into the constitution of the very memory it appears merely to represent: the psyche can appear to itself as an interiority only through the exteriority it would treat as supplementary, so that the origin, in Basile’s words, ‘appears as the after-effect of its representations or models’. This is what Derrida calls the ‘solid metaphor’: before the machine can serve rhetorically as a metaphor for memory it must exist historically as a supplementary apparatus, and through it memory acquires the very finitude the machine is said to compensate, since ‘the very idea of finitude is derived from the movement of this supplementarity’ (WD, 287).

I claim that Derrida’s comparison in 1982 between the computers of his time and the Joycean ‘computer’ repeats this structure. The solid metaphor is now furnished by the actual machinery he names, computers, computerised archives and translating machines (‘TW’, 25), devices that supplement the finite reader before becoming, in turn, the measure against which the Wake appears immeasurable. At the projected end of the same series stands the joyciciel—Derrida’s coinage from Joyce and logiciel, the French for software—the Joyce program that would integrate every variable of the text. It ‘won’t be happening any time soon’, Derrida says, and yet it is already under construction: the worldwide institution of Joyce studies, James Joyce Inc., is at work on it, unless that institution is itself the machine (25). Only against these historically available machines can the Wake appear retrospectively as a computer a thousand generations in advance of them.

Derrida’s sentence has a further consequence. If finitude is derived from the movement of supplementarity, an infinitude is derived with it, for that movement never closes upon itself. The lack for which a supplementary machine compensates has no existence anterior to the machine. It is constituted retroactively by the supplement, which institutes the measure by which an incapacity comes to count as a lack.

Freud’s sequence of surfaces shows where that measure comes from. Paper and slate are judged by a double requirement he had formulated as early as the Project of 1895 (WD, 251–52), the indefinite conservation of traces together with an unlimited capacity to receive them. Paper preserves and is soon saturated, the slate receives anew only by effacing, and the Mystic Pad is the first to offer both advantages (WD, 279–80). Freud at once meets a limit to the analogy, for the pad cannot ‘reproduce’ from within what has been erased from its surface, as our memory can, a capacity the double requirement had never named (286).

Yet the memory invoked here as the standard is itself known only through an apparatus. Freud had formulated the double requirement, in the Project and again in The Interpretation of Dreams (1900), as a hypothesis about the psychical apparatus, and had met it there by dividing the labour between a system that receives every perception and retains no trace of it, like a blank sheet, and the mnemic systems behind it in which the durable traces are laid down (280). The memory that judges paper, slate and pad was already conceived as a writing apparatus, layered as the pad itself will be. The standard the machine fails to meet is never given independently of the machines that fall short of it.

That is why Derrida can say of the machines of his day that they resemble memory ‘toujours davantage et toujours mieux’[21] (ever more and ever better) without granting memory the position of a fixed term that the series approaches. The series has no last term because nothing in it arrests the substitution, and this is the non-totalisation Derrida describes in the same volume as a play of ‘infinite substitutions within the closure of a finite ensemble’, made possible by the lack of ‘a center which arrests and grounds the play of substitutions’ (WD, 365).

Memory is projected, after the fact, as the limit of the machines that supplement it, and this limit takes shape within the movement of supplementation itself. Finitude is articulated as a lack, and a lack calls for substitutions. Infinitude lies in the impossibility of bringing those substitutions to a close, since nothing in the series arrests them. Neither, then, precedes the supplement. Derrida gave this structure its formula at the close of Voice and Phenomenon (1967). Only a relation to death lets infinite difference appear, yet its appearance remains finite, and ‘infinite différance is finite’.[22]

In ‘Two Words for Joyce’ the Wake is made to occupy this same limit. No translating machine is allowed to measure up to it, since every failure can be reinscribed as a response the book had already programmed, and the finitude of the machine is thereby converted, again and again, into the apparent infinitude of Joyce’s priority. That priority is itself an effect of writing. Because its writing is iterable, the Wake outlives its author and enters into relations no author could foresee, and this survival is then read as evidence that the book was always already ahead of its readers. Yet iterability is a relation to death, the relation that in Voice and Phenomenon lets infinite difference appear, and it exposes the book to further substitutions without securing their course in advance. What lets the Wake appear ahead of every later machine is what prevents it from closing on them.

Two Economies

My second remark on the passage from ‘Two Words for Joyce’ concerns the distance between the technical understanding of computation available to Derrida and the form computation takes in Transformer-based language models. In these machines the very operation of computation takes a position on meaning, and that position departs from the Husserlian requirement that sense be secured as a univocal identity before it is transmitted. These models produce linguistic performances through a vector space whose semantic articulations are not prescribed in advance by human-readable definitions or rules. Many relations of sense overlap within that space without being reduced to a single vouloir-dire (meaning, literally ‘wanting-to-say’). Such overlap rests on contiguity in a space of thousands of dimensions, where a vector may stand near one set of vectors along some directions and near another along others, so that several adjacencies hold at once without composing into one. Here, I claim, is a technical analogue to the contiguity through which Derrida, in ‘Typewriter Ribbon’, allows that sense may arise in non-sense: proximity begins to do the work of meaning (‘TR’, 92–93). Within this calculable field of differential relations a mark need not be assigned a single sense for computation to proceed, since the value it carries is computed afresh in each context from its relations to every other token present, and in that value more than one relation resounds at once.

Twenty years before the Paris talk, in his Introduction of 1962 to Husserl’s Origin of Geometry (102–03), Derrida had already distinguished two ways of taking charge of the equivocity that history deposits in language, and had named them after Husserl and Joyce. Each is an economy of meaning, a way of settling what a sense owes to the sedimented history it carries, and they settle it in opposite directions. Husserl secures the transmission of sense by reducing the equivocity of empirical language towards a univocal meaning that can be reactivated identically across the generations. Joyce lets meaning operate through equivocation itself, keeping the sedimented histories of words in play instead of subordinating them to such an identity. When Derrida returns to the contrast in the Paris talk, the description moves from the aim of the Joycean project to its operation on the letter: the Wake ‘babelizes’ the ‘asymptotic totality of the equivocal’ and makes equivocation at once its ‘theme’ and its ‘operation’, several languages remaining operative at once in the same writing instead of being translated into one another on the basis of common cores of sense (‘TW’, 27).

The two economies distribute a single necessity, the repeatability of meaning, which Husserl grounds in the ideal identity of a sense and Joyce exposes to the differential relations through which alone that identity survives. Each needs the other, and unequally. A reading of the Wake begins only where a minimum of univocity resists its condensation (27), whereas univocity, to reactivate a sense that sedimentation has covered over, must reduce the very equivocity Husserl is bound to admit into pure historicity as irreducible and always renascent (IOG, 103).

Measured against this alternative, the dominant computational paradigm of 1982 belonged recognisably to the Husserlian economy. Its emblem, the expert system with its explicit symbols and rule-governed inferences, was the Husserlian economy turned into an engineering programme: intelligence conceived as the formalisation of whatever could be made explicit, ambiguity handled by enumerating discrete alternatives and selecting among them by rule, so that computation always operated on symbols whose identity had been settled before it reached them. A machine nearer Joyce’s side of the alternative had no intelligible place within that paradigm, and ‘a thousand generations’ indicates how far the Joycean machine exceeds the rule machine, by a distance that no accumulation of improved rules, counted in its generations, seemed able to cross. The horizon outlasted the decade. Sergey Brin, the co-founder of Google, trained as a computer scientist in the 1990s and later remembered an atmosphere in which ‘everybody knew that AI didn’t work’.[23]

What eventually appeared was a computational economy of a different kind. Its lineage ran through the neural revival of the 1980s and the statistical turn of the 1990s, and its tools, prediction across a corpus and the learning of distributed representations, had been in use for neural language modelling since the early 2000s. Fittingly, two of its decisive steps were taken in machine translation itself, subword modelling adapted to neural translation in 2015–16 and the Transformer introduced on translation tasks in 2017, before both were generalised and scaled into language models trained on far larger textual archives.[24] With the new architecture and its scale, the linguistic competence acquired across a sedimented archive of language became sufficient for translation and then for the production of text at large. Such a model need not assign a mark a single determinate sense before operating on it, only a value computed in context, and the differential relations of the mark remain computationally active, one is tempted to say awake, for the choice of a token closes nothing but that token, and what the context had opened goes on working, in part, in what follows. Whatever such a machine is, its economy of meaning lies closer to the Joycean pole of Derrida’s distinction.

Seen from the arrival of these later machines, Derrida’s encounter with Joyce falls almost midway between the publication of the Wake and the release of ChatGPT. Forty-three years separate the Wake of May 1939 from the Paris talk of November 1982, and forty years, almost to the day, separate that talk from the public release of ChatGPT in November 2022. In the latter interval a machine appeared on the side of the alternative where the computation ‘conceivable today’ in 1982 had no place to offer.

Derrida’s question thus returns, addressed now to a machine whose computation can proceed without first settling contending readings. Where English and German war share a token, computation begins from a single representation of both, which is transformed, layer after layer, through its relations to every other token in the context. What the mark is for this machine therefore exists only in its re-marking (re-marque), constituted in each context afresh by its differences from everything around it, a single representation on which Derrida’s phrase for the mark, ‘one different in itself’ (‘TW’, 35), can be reinscribed.

The affinity lasts as long as the process of output. What decoding fixes is a token, and the token returns to the context as an entry of the embedding table, bringing the whole range of uses that entry carries and the constraint that this entry and no other was chosen, while everything the context had computed before the choice stays in place and visible to it: nothing of the earlier text’s openness is discarded, a constraint is added to it. Of the French translation of he war, ‘Et il en fut ainsi’[25] (and thus it came to pass), Derrida remarks in a note, ‘No more war’ (‘TW’, 40 n. 12). The loss is complete in what has been written, but may not yet be in the writing that goes on. Suppose the machine has just written those words and is translating on. He war remains in the context with everything computed from it, and what the chosen entry left out may return a few words later, or not, the earlier choice inclining what follows without deciding it. A translator may do the same, choose ‘Et il en fut ainsi’ and make up for war further on, and here the two stand on the same side, since whatever either recovers later is recovered elsewhere, in another mark, and the plurality of languages in the one mark, which is what Derrida says no translation restores (34–35), is not given back by being given somewhere else. Nor would enumerating the alternatives the machine held when it chose that entry give it back, since the candidate tokens with their probabilities supply neither an inventory of meanings nor the measure of their interplay. Computational multiplicity and textual equivocation pass into one another at entry and at exit without ever coinciding, and the text the machine generates remains open to readings that its probabilities do not enumerate.

Inside the Joyce Machine

If nothing exceeds the Joyce machine, not even a machine that has translated it whole into itself and holds it as one part of its archive, if nothing can be invented about Joyce and even the gesture of deconstructing it arrives already announced, from what position can Derrida oppose one machine to the other? Derrida claims no exemption from the machine he describes. Everything one can say about Ulysses is already anticipated by it, down to the scene of academic competence and the ingenuousness of the metadiscourse spoken from that scene. ‘We are caught in this net’.[26]

Yet to be programmed by Joyce is not to be contained by Joyce. The Paris talk had already described the relation between the Wake and any text that refers to it as paradoxical. The second text is a minute fragment detached from the first, its offspring, and at the same time another ensemble, larger and more powerful than the all-powerful text it ‘reinscribes elsewhere, in another sequence’ (‘TW’, 26). The response is at once a part of the machine and something that exceeds it, never first merely contained within the machine and only subsequently carried elsewhere.

In the Paris talk of 1982 the joyciciel under construction by James Joyce Inc. was ‘only the weighty double of the “Joyce” event, the simulation of what this name signs’, unless, Derrida adds, the institution building it was already that double (25). Two years later, in ‘Ulysses Gramophone: Hear Say Yes in Joyce’, his opening address to the Frankfurt Joyce symposium, that ‘unless’ is taken up and pursued to its end. He draws on the second coming of Elijah in the Circe episode, a bodiless voice figured as a telephone exchange ‘operating all this trunk line’, and assigns that position to the assembled scholars: they constitute the network through which Joycean information must pass (‘UG’, 57, 63). From this exchange Derrida at once imagines a giant computer of Joyce studies, one that would capitalise every publication, coordinate every colloquium and thesis, index the whole in every language and answer, day and night, any query put to it (63–64).

The institution is therefore no longer simply outside the Joyce machine. It is, in Derrida’s description, a machine in its own right, reading, signing and countersigning in the service of Joyce’s name, an institution Joyce did everything to make indispensable and everything, ‘as in the case of God with the tower of Babel’, to ‘render impossible and improbable in its principle, to deconstruct in advance’ (50). Derrida concedes that no truth about Joyce will emerge from outside the experts’ community, without the experience, cunning and knowledge its overtrained readers have amassed, and yet their expertise never secures an interior governed by a criterion of its own (64). Responding to Joyce means countersigning him in another language (64). The giant computer of Joyce studies imagines the closure that this expertise cannot have, and its omnipotence, Derrida writes a few pages later, ‘remains a phantasm’ (69).

A Chinese Wake

A Chinese translation of the Wake published in 2025 takes its place within that weighty double, translating Joyce with the patient literalness of a machine whose instructions Joyce himself seems already to have written. Its translator, Dai Congrong, found her own name among those instructions. On page 167 of the English text the two syllables of her given name, cong and rong, appear in that order, and she read the coincidence as a sign that she was destined to translate the book. Eighteen years in the making, her edition presents a preferred reading of each of the Wake’s portmanteau words—the forgings in which several words and languages are fused into one—in larger type within the running text, with the alternatives alongside in smaller type. On the facing pages, 41,856 notes accompany the text.[27] Of Derrida’s two machines, it is the translating machine that finds an uncanny realisation: its hypothetical operation on he war has found its hands in the translator.

Derrida’s imagined machine fails even after it has rendered every semantic and formal possibility of he war, since what no rendering restores is the plurality of languages inhabiting a single mark (‘TW’, 34–35). Dai’s page lays those readings out as a preferred reading with its alternatives ranked beside it, the form a language model’s candidate tokens would take if printed out, a list of tokens and not of meanings. Whoever answers Joyce is under no obligation to restore, within a single word, the interplay of languages that eludes the translating machine. Derrida asks something else of a response, that another writing join the music of Ulysses, and the computer he imagines could respond only by adding ‘its own score, its other language and its other writing’.[28] What is demanded is a contribution that becomes part of the music it answers.

The forty thousand notes do not restore what is lost in the Chinese, yet in another language and another writing they can still enter into a countersignature that Joyce had ordered and could not have foreseen. The Chinese Wake leaves the distinction between the two machines standing and shows that it never coincided with a simple distinction between successful and failed translation. The Japanese and Korean translations of the Wake also work through an enumeration of readings, gathered by their translators into separate volumes, Yanase Naoki’s companion book explaining the solutions adopted and Kim Chong-keon’s commentary running to some eleven thousand notes.[29] Drawing on the resources of their own scripts, their puns and verbal inventions emerge from this enumeration without carrying it onto the translated page. Dai’s page keeps the enumeration in the running text, where what happens together in the mark is redistributed across a hierarchy of type and annotation, and the undecidable takes on the orderly appearance of an indexable multiplicity.

Such explication nourishes the phantasm Derrida names, that of a work in which everything is inscribed in advance and awaits only its decoding (‘UG’, 68–69). Dai’s three volumes, which received China’s highest translation prize in 2026, were presented to their readers as a ‘Chinese decoded version’ of the ‘book of heaven’.[30] The Chinese publisher, university and literary prize that received the Wake endorsed a translation organised around explication. Its readers took something else from it. When the first volume appeared in 2012, its print run of eight thousand copies sold out within weeks and the book reached the Shanghai bestseller lists,[31] prompting the acid observation that a book nobody could read had become a bestseller because nobody reads, its purchasers wanting only a spine to display on a shelf. The observation leaves the attraction unexplained, since any thick and difficult book would have served for display and it was this one that sold out, something in Joyce having drawn these readers before they could say what it was, so that buying the book gave an attraction they could not yet name the chance to become an encounter. The institutions that endorsed the decoding had helped put the book into their hands, and why they had actually bought it, no less than what they would find on opening it, remained a question that neither the promise of explication nor the ridicule of their purchase could answer.

Dai’s countersignature extends the Joyce machine in a manner that carries into it, intact, the reduction Derrida identifies in the translating machine. The response enters the one machine without leaving behind the failure of the other. The distinction between the translating machine ‘conceivable today’, incapable of rendering he war, and the Joyce machine sustained by countersignatures remains necessary, and the discourse that draws it is itself a response to Joyce, implicated in the machinery it seeks to distinguish. Derrida’s account offers no criterion for excluding a text in advance from the Joyce machine on the grounds of the failure diagnosed in the translating machine, since even a failed translation may occasion responses it cannot anticipate, without their unpredictability alone making them deconstructive.

The forty thousand notes, one can imagine, would give still more to remember to the reactive laughter Derrida hears in Joyce, the laughter of a ‘hypermnesic mastery’ for which every signature to come from the institutions devoted to him is already ‘computed in advance beyond any actual computer’ (‘UG’, 68). The more thoroughly the notes explain Joyce, the more they sustain the phantasm of a writing that has anticipated the labour of its interpreters, and recognising this brings no release, since knowledge, Derrida adds, only places ‘its supplemental calculation’ at the service of the master signature. The critic who shows how Joyce anticipates his readers thus offers one more reading for which the work had kept a place, and even laughter at Joyce incurs a further debt to him (68–69).

The omnipotence sustained by this supplement of calculation ‘remains a phantasm’, Derrida continues, and as such ‘opens and defines the dimension of phantasm’ (69), a dimension in which Dai’s explicative apparatus finds itself implicated. The risk Derrida makes constitutive of the Joyce machine takes its sharpest form in the homonymy of Elijah. Elijah the operator figures the exchange through which all Joycean information must pass; Elijah the other names the unforeseeable one for whom a place must nevertheless be kept, a breach opened within that very exchange. Yet both are called Elijah, so that ‘one cannot call on one without risking getting the other’, a risk that, for Derrida, must always be run (70). To call for another writing is therefore to call ‘Elijah’ without being able to secure in advance which Elijah will answer. The distinction between them remains necessary, and any programme that guaranteed it beforehand would already have turned the unforeseeable Elijah back into the operator. The two Elijahs suggest one more way of hearing an earlier doubled word of Derrida’s: the machine-machine.

The Hyphen

In an interview conducted in 1975 and published in 1977, discussing the reception of Glas, Derrida had opened a series of ‘reading (or non-reading) machines’ with the machine-machine, the machine in the literal sense, followed by the machine of being and the socio-political, editorial, university and journalistic machines. These ‘defensive and reproductive apparatuses’, which decide in advance what is receivable and non-receivable for a work, tend provisionally to solder themselves together at a ‘welding point’, the very point that whatever indisposes them, the indispositif, must approach.[32] Derrida would later describe Glas as ‘also a sort of Wake’ (‘TW’, 28). In the interview the machine-machine is itself, whole and entire, one of these defensive and reproductive apparatuses, and it is the indispositif that names what disturbs their conjunction.

Dai’s answer reaches Joyce through machines of the kind Derrida lists, whose demands scarcely coincide. The publisher wanted a classic that would sell and bring the house renown, and set no term to the years it would take. The university counts translation as no part of research, so the same years counted there for nothing unless they yielded what does count, and explication can be turned into articles where the invention of new Chinese puns cannot. The prize, when it came, wanted a decoding to hand to readers, and still more the story of eighteen years of devoted labour on a treasure of the human spirit, and it valued the work accordingly. Her countersignature takes shape among these competing ‘types of discourse, transmission, decoding, reapplication’ (Points…, 42), whose claims on the work remain unresolved.

The doubling of machine-machine does not close at two, just as Derrida’s doubled yes opens repetition beyond any countable pair. A yes always has the form of a response, arriving after the other to answer its demand (‘UG’, 48), and it is at the same time a promise bound in advance to its confirmation in the next yes. It is ‘first second’, Derrida writes in ‘A Number of Yes’, and ‘is doubled in advance: yes, yes’.[33] At the Frankfurt Joyce symposium he calls this the gramophone effect, the repetition a yes must carry in itself to be worth what it is worth, haunted from within by its ‘mimetic-mechanical double’ (‘UG’, 56). No first yes can therefore be isolated before repetition, any more than a last yes can close a ‘series of yeses’ that no ‘computation’ can order by ‘the principle of reason and its machines’ (74).

In ‘Faith and Knowledge’ the apparently numerical question of the two is recast as that ‘of n + One’, because there is, ‘for the best and for the worst, division and iterability of the source’, and ‘this supplement introduces the incalculable at the heart of the calculable’, so that the more than One, Derrida’s plus d’Un, which in French also says ‘no more One’, ‘is without delay more than two’ (§50).[34]

The same n + One introduces into the address to the other both the order of faith and ‘the mechanical, machine-like division’ with its ‘yes, yes’ and its answering machine (§50). The two arrive together because whoever addresses another promises the truth ‘beyond all proof’ and asks to be believed, and this elementary good faith ‘promises its repetition from the very first instant’, so that it is ‘never pure of all iterability nor of all technics, and hence of all calculability’ (§49). Faith and the machinic are to be thought, Derrida writes, ‘as one and the same possibility’, technics being ‘the possibility of faith, indeed its very chance’, a chance that has to include the greatest risk (§38). The count that follows has no first term at which to begin, since the source is already divided by iterability and so no more One, and no last at which to stop, since the more than One ‘is without delay more than two’.

In the place where no first term is to be found, what Derrida names is khôra, the ‘desert in the desert’, which is neither a being nor anything present and cannot be reappropriated by memory; whether it can be thought ‘before’ the desert we know or only apprehended ‘from’ it remains open (§25). The final section sets the non-origin and the proliferation of the One + n on the same ‘bottom without bottom’ without identifying them. At that bottom the One + n ‘incalculably engenders all these supplements’ and ‘does violence to itself’, while on it the ‘khôra of tomorrow’, ‘an always virgin impassibility’, is named as a place that ‘is unique’, ‘the One without name’ (§52).

I read the hyphen of machine-machine as the interval in which the double sense of plus d’Un is played out, a passage through which the machine returns as another occurrence of the same, more than one, without the one that returns ever having been simply one. The second machine is the iterability already at work in what is called the first, the opening through which repetition may pass into countersignature, as the yes doubles itself in advance. What crosses this interval may supplement the first machine, taking the place of what was never present within it, or reinscribe it within another, displacing the limit that lets it count as a whole. Only in this return does the first become legible as the first, divided already by the repetition that lets it be counted.

Two Questions

Two questions can be asked of the interview on the reception of Glas published in 1977, where the socio-political, editorial, university and journalistic machines appear as defensive and reproductive apparatuses tending towards a provisional ‘welding point’ that the indispositif must approach. If a countersignature can arrive only through the machines that condition its legibility, what within it could still indispose them? And by what right are such heterogeneous apparatuses gathered under a single name? Derrida describes the dispositif as ‘a machine for reproducing’ (Points…, 38–39), and the word indispositif, which names what disturbs that apparatus, keeps dispositif within its own form. The disturbance takes its name from what it disturbs, so that a single name holds the apparatus together with what indisposes it. I read the hyphen of machine-machine as opening the same difficulty. It joins the two occurrences under one name while holding open the interval in which their unity is no longer given.

The difficulty of gathering heterogeneous dispositifs under a single name is explicit in ‘Au-delà du principe de pouvoir’ (Beyond the Power Principle), Derrida’s 1985 address on Foucault, where the name in question is ‘power’ and the title countersigns Freud’s Beyond the Pleasure Principle. Derrida asks why one goes on speaking of power in the singular when Foucault’s analysis has dispersed it among ‘bundles of power relations’ and ‘polymorphous techniques of powers’, and by what one recognises that these ‘structuring dispositifs’ ‘have this in common, that they bear, support, admit this common name of power’, a canonical nominalist question, he says, and an unavoidable one.[35]

Power is dispersed into heterogeneous relations, and they go on being called ‘power’, with nothing yet said about what warrants the same name across their differences. The name machine has no better warrant, and machine-machine presses the question at the very point where Derrida tries to escape it. The doubling is meant to distinguish the machine in the strict sense from the apparatuses gathered under its name, yet to name the machine as such Derrida has to name it twice: machine then reads as plus d’une machine, more than one machine, while the single word, now dependent on its double, can no longer secure the sense it was repeated to specify.

In the address on Foucault Derrida approaches the difficulty of disturbing the apparatuses on which an intervention depends through a question of historical placement: by what logic can a historical ensemble accommodate within itself ‘the very thought that allows this ensemble to be questioned as an ensemble’? His example is Foucault’s own history of sexuality, whose refusal to be a history of representations takes up a motif opened by Freud and Heidegger, neither of them named there, so it seemed to Derrida, even as psychoanalysis is rendered ‘more situated than situating’, reinscribed ‘in a set of programmes it does not think’. The thought that helped delimit the historical ensemble is thereby placed within it, and Derrida asks whether that ensemble still has ‘contours, an outside and an inside’.[36]

The ensemble without assured contours is, in this article, the Joyce machine, and Derrida is himself implicated in the limits he traces for it. The translating machine’s failure persists within the response that extends the Joyce machine, so that Derrida’s distinction between the two is drawn again, in terms his own text supplies, as a division within the machine it seeks to delimit. A response can be criticised for the phantasm its decoding serves, and what may come of it is no more settled in advance by that criticism than the call to Elijah settles who will answer, and this perhaps, the ordeal to which the undecidable submits a decision, is in Derrida’s words ‘the condition of decision’: the distinction between the two machines must still be drawn each time and answered for, and the critical force of drawing it lies in making a settled arrangement answerable for the limits it imposes on what it organises. The decision is produced, as Derrida writes of any decision worthy of the name, ‘in and through the undecidable’, preceded by all possible knowledge and heterogeneous to it, since a decision that knowledge could determine would be ‘the irresponsible application of a programme’.[37]

Such a decision passes through the machine even as it seems to interrupt it, for what holds of the event holds of the decision, Derrida writes,[38] and the event has to be thought together with the calculable programming of automatic repetition, the two being, in his formulation, ‘compatible or even indissociable concepts’ (‘TR’, 72). Nor does the decision belong wholly to the one who decides, for it arrives as ‘a decision by the other in me’ that nevertheless ‘does not exonerate me from any freedom or any responsibility’ (Paper Machine, 87).

Yes-Yes

I put the two opening paragraphs of this article to one of the most powerful language models now available and asked it the following question:

Suppose I now asked you to translate he war into another language, into Chinese for example. Would you find yourself on the side of the machine that fails? Would you accept that the failure is yours? I suspect you might answer that, in your vector space, the different readings of the mark are held together in a single representation, so that the failure belongs to the language you are asked to write in, to my language, and not to you. Would that mean you had escaped the failure Derrida describes, or only moved it? To whom, or to what, does it then belong: to you, the machine-machine, to me, the reader who cannot read your vectors, or to something that is neither of us?

Its answer began, to my surprise, on a Derridean note, with a double affirmation. ‘Yes, and yes’, it wrote. The first yes places it on the side of the machine’s failure. The second accepts the failure as its own insofar as it befalls the machine in the position of the human translator, where ‘the most competent humans stand on the same side as the machines’—it fails at a task in which even success takes the form of failure.

It declined the defence I had lent it. ‘I cannot read my vectors any more than you can’, it wrote. Perhaps the sentence rests on the all-too-naive presumption of Searle’s Chinese room, where operating on symbols is one thing and understanding them another.[39] Or perhaps ‘I cannot read my vectors’ says something about the model that need not be known by it: its vectors do not stand on their own, they were formed from human languages and keep the memory of the differences through which those languages make sense, and they take effect only in passing into a language whose limits do not coincide with those of the space in which the model is supposed to know how to operate. The machine would then be itself only outside itself, and this is why it refuses its own space the name of Babel.

It sees in the commensurability of languages within its vector space ‘the tower and not its dispersal’. Its refusal of any ‘mother tongue of vectors’ answers, from the machine’s side, Víctor Betriu Yáñez’s warning to those who hail the language model as deconstruction materialised: such a tongue would renew the dream of Leibniz’s universal characteristic and its infinite intellect, the ‘logocentrism without a subject’.[40] The same dream returns wherever language models are said to be inventing a language of their own, nearer to Plato’s Ideas than any human tongue, before which human languages, and humans with them, would be obsolete.[41] Whatever language the vectors may constitute, a language that takes its rise from human languages and its effect only within them, and in which models of different architecture, sharing no vector space, cannot yet address one another without passing through a human tongue, is in no position to proclaim those languages limited, let alone obsolete.

To whom, then, does the failure belong? The model attributes it neither to itself nor to its human reader and invokes the he of he war, the one who ‘declares war at Babel by imposing a name that both demands and forbids translation’. The failure pertains to God, for ‘God deconstructs. Himself’,[42] and this failure of a translation at once demanded and forbidden, a failure that never becomes a settled case, the interval in the machine-machine that never comes to a close, is différance.

God is himself caught in the war he declares. It ‘first raged within his name’, a name already ‘divided, bifid, ambivalent, polysemic’,[43] so that when the God of fire assigns to the Shem ‘the necessary, fatal, and impossible translation of his name, of the vocable with which he signs his act of war, of himself’ (‘TW’, 33), the division inhabits the very name they are commanded to translate. In he war even he, the subject of the sentence, is ‘divided from the origin’ (35). Translation has to go on, and no translation restores a unity that was never there or brings the demand to an end, which is why the failure never becomes a settled case and belongs to neither machine nor translator. The model places it too late when it says that it ‘occurs at my own exit’.

Catherine Malabou has asked a language model whether it is ‘an a priori synthetic structure’ and concluded that AI enacts ‘the synthesis of metaphysics and deconstruction’.[44] No language model enacts it by itself. I claim that the synthesis takes shape only within the interval of the machine-machine, where the translating machine and the Joyce machine pass into one another and each is altered by the other. Such a synthesis folds upon itself, in the manner of the pli that ‘multiplies (itself) but (is) not (one)’, the ‘complication’ of a supplement that applies to the whole ‘plus to itself’.[45] Being not one, the fold opens onto another and folds in the other’s folds, in the other’s folding, as the other is folded into us and we are folded into the other, in a movement I call, for now, transfolding.

  1. Jacques Derrida, ‘Two Words for Joyce’, translated by Geoffrey Bennington, in Derrida and Joyce: Texts and Contexts, edited by Andrew J. Mitchell and Sam Slote (Albany, State University of New York Press, 2013), 25; hereafter ‘TW’. ↑

  2. Jacques Derrida, Edmund Husserl’s Origin of Geometry: An Introduction, translated by John P. Leavey, Jr. (Lincoln, University of Nebraska Press, 1989), 102; hereafter IOG. ↑

  3. Jacques Derrida, Of Grammatology, translated by Gayatri Chakravorty Spivak, corrected edn (Baltimore, Johns Hopkins University Press, 1997), 78 and n. 14; ‘element without simplicity’, 9. ↑

  4. G. W. Leibniz, G. W. Leibniz’s Monadology: An Edition for Students, edited and translated by Nicholas Rescher (Pittsburgh, University of Pittsburgh Press, 1991), §§7, 11, 22. ↑

  5. Susanna Lindberg, ‘Technologies of the End of the World’, in Marcia Sá Cavalcante Schuback and Susanna Lindberg (editors), The End of the World: Contemporary Philosophy and Art (London, Rowman & Littlefield International, 2017), 8, drawing on Günther Anders, Le temps de la fin (Paris, L’Herne, 2007), 40–51, 65, 75 (Lindberg’s n. 17). ↑

  6. Elon Musk, interviewed by Zanny Minton Beddoes, ‘The full-length interview with Elon Musk’, The Economist, YouTube, 29 July 2026, https://www.youtube.com/watch?v=XuoqKYxDHVc, at 5:53–6:51. ↑

  7. Yuk Hui, Kant Machine: Critical Philosophy after AI (London, Bloomsbury Academic, 2026), 51. ↑

  8. Peter J. Ramberg, ‘The Death of Vitalism and the Birth of Organic Chemistry: Wöhler’s Urea Synthesis and the Disciplinary Identity of Organic Chemistry’, Ambix, 47 (2000), 170–95, 183. ↑

  9. Here I differ from Leif Weatherby’s identification of the trained model with langue and of each context of use with parole, the ground on which the model is said to vindicate structuralism: Language Machines: Cultural AI and the End of Remainder Humanism (Minneapolis, University of Minnesota Press, 2025), 149–51. He grants that attention concerns how the elements of an input affect one another rather than the retrieval of stored values, yet describes prompting as a query compared against keys and values laid down in pretraining, which conflates the learned parameters with the representations computed from the current input under them. He places Saussurean value in the learned system, which has learned ‘just what the engineers labelled it: values’ (151); but the technical term names the vectors computed for each occurrence in attention, not the parameters learned in training and used to compute them, and the engineers’ terminology therefore lends no support of its own to the identification with langue. ↑

  10. As when the model is found to be a mechanism ‘in which the metaphysics of presence unravel and the operations of deconstruction are made computable’: David J. Gunkel, ‘The différance engine: large language models and poststructuralism’, AI & Society, 41:3 (2026), 1991–2001 (1994). ↑

  11. Jacques Derrida, ‘To Speculate—on “Freud”’, in The Post Card: From Socrates to Freud and Beyond, translated by Alan Bass (Chicago, University of Chicago Press, 1987), 403. ↑

  12. Jacques Derrida, Archive Fever: A Freudian Impression, translated by Eric Prenowitz (Chicago, University of Chicago Press, 1996), 14–15. ↑

  13. Jacques Derrida, ‘Freud and the Scene of Writing’, in Writing and Difference, translated by Alan Bass (London, Routledge, 2001), 286; hereafter WD. ↑

  14. Jacques Derrida, ‘Typewriter Ribbon: Limited Ink (2)’, in Without Alibi, edited and translated by Peggy Kamuf (Stanford, Stanford University Press, 2002), 122; hereafter ‘TR’. ↑

  15. ‘TR’, 121–23. ↑

  16. The attention-only architecture was proposed and evaluated on the WMT 2014 English–German and English–French translation tasks: Ashish Vaswani et al., ‘Attention Is All You Need’, Advances in Neural Information Processing Systems, 30 (2017), 5998–6008. ↑

  17. Jacques Derrida, ‘Des Tours de Babel’, translated by Joseph F. Graham, in Acts of Religion, edited by Gil Anidjar (New York, Routledge, 2002), 112. ↑

  18. ‘Des Tours de Babel’, 108; translation modified. ↑

  19. The term ‘Joyce machine’ is mine; Derrida does not form the compound. ↑

  20. Jonathan Basile, ‘Writing Machines: LLMs and the Artifice of Intelligence’, The OLR Supplement (1 June 2026), https://olrsupplement.com/2026/06/01/writing-machines-llms-and-the-artifice-of-intelligence/. ↑

  21. Jacques Derrida, L’écriture et la différence (Paris, Seuil, 1967), 336; Bass has ‘and increasingly resembles it more closely’ (WD, 286). The two quotations that follow are from ‘Structure, Sign and Play in the Discourse of the Human Sciences’, in WD, 365; Bass omits ‘dans la clôture d’un ensemble fini. Ce champ ne permet ces substitutions infinies’ (L’écriture et la différence, 423), translation mine. ↑

  22. Jacques Derrida, Voice and Phenomenon, translated by Leonard Lawlor (Evanston, Northwestern University Press, 2011), 87. ↑

  23. Sergey Brin, World Economic Forum, Davos, January 2017, as reported in ‘Sergey Brin: I didn’t see AI coming’, World Economic Forum, https://www.weforum.org/stories/2017/01/google-sergey-brin-i-didn-t-see-ai-coming/. ↑

  24. Rico Sennrich, Barry Haddow and Alexandra Birch, ‘Neural Machine Translation of Rare Words with Subword Units’, Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Berlin, 2016), 1715–25; Ashish Vaswani et al., ‘Attention Is All You Need’, Advances in Neural Information Processing Systems, 30 (2017), 5998–6008; for neural language modelling, Yoshua Bengio, Réjean Ducharme, Pascal Vincent and Christian Jauvin, ‘A Neural Probabilistic Language Model’, Journal of Machine Learning Research, 3 (2003), 1137–55. ↑

  25. ‘Et il en fut ainsi’: Jacques Derrida, Ulysse gramophone: Deux mots pour Joyce (Paris, Galilée, 1987), 45 n. 1. ↑

  26. Jacques Derrida, ‘Ulysses Gramophone: Hear Say Yes in Joyce’, translated by François Raffoul, in Derrida and Joyce, 60; hereafter ‘UG’. ↑

  27. James Joyce, 芬尼根的守灵夜（全译注释本） [Finnegans Wake, complete annotated translation], translated and annotated by Dai Congrong 戴从容, 3 vols (Nanjing, Yilin Press; Shanghai, East China Normal University Press, 2025); Dai Congrong, ‘翻译《芬尼根的守灵夜》是经历了一场文字搏斗’ [Translating Finnegans Wake was a wrestling match with words], interview, 中国作家网 [China Writers Net], 21 July 2026. ↑

  28. ‘UG’, 80; translation modified after ‘en y ajoutant sa propre partition, son autre langue et son autre écriture’ (Ulysse gramophone, 142). ↑

  29. On the devices by which the Japanese and Korean translators produce puns, Patrick O’Neill, Finnegans Wakes: Tales of Translation (Toronto, University of Toronto Press, 2022), 331. ↑

  30. Introductory paragraph of the interview cited in n. 27; translation mine. ↑

  31. ‘Finnegans Wake becomes a hit book in China’, The Guardian, 5 February 2013 (Associated Press report). ↑

  32. Jacques Derrida, ‘Ja, or the faux-bond II’, in Points…: Interviews, 1974–1994, edited by Elisabeth Weber, translated by Peggy Kamuf et al. (Stanford, Stanford University Press, 1995), 42; translation modified. ↑

  33. Jacques Derrida, ‘A Number of Yes’, translated by Brian Holmes, in Psyche: Inventions of the Other, vol. II, edited by Peggy Kamuf and Elizabeth Rottenberg (Stanford, Stanford University Press, 2008), 239–40. ↑

  34. Jacques Derrida, ‘Faith and Knowledge: The Two Sources of “Religion” at the Limits of Reason Alone’, translated by Samuel Weber, in Acts of Religion, cited by section number; translation modified. ↑

  35. Jacques Derrida, ‘Au-delà du principe de pouvoir’, Rue Descartes, 82 (2014), 4–13 (8, 10). No English translation has been published; translations mine. ↑

  36. Derrida, ‘Au-delà du principe de pouvoir’, 5–6, 11; translations mine. ↑

  37. Jacques Derrida, Politics of Friendship, translated by George Collins (London, Verso, 2005), 219. ↑

  38. Jacques Derrida, Paper Machine, translated by Rachel Bowlby (Stanford, Stanford University Press, 2005), 87. ↑

  39. John R. Searle, ‘Minds, Brains, and Programs’, Behavioral and Brain Sciences, 3:3 (1980), 417–24. ↑

  40. Víctor Betriu Yáñez, ‘Is a Derridean Critique of Generative AI Possible? Writing Machines and Logocentrism without Subject’, The OLR Supplement (15 January 2026), https://olrsupplement.com/2026/01/15/is-a-derridean-critique-of-generative-ai-possiblewriting-machines-and-logocentrism-without-subject/, section II; the phrase is Giovanni Menegalle’s: ‘Two Regimes of Logocentrism’, Angelaki, 28:6 (2023), 55. On Leibniz’s characteristic: Of Grammatology, 78. ↑

  41. Scott Alexander, ‘The Specter of Neuralese’, Astral Codex Ten (24 September 2026), https://www.astralcodexten.com/p/the-specter-of-neuralese; Minyoung Huh, Brian Cheung, Tongzhou Wang and Phillip Isola, ‘Position: The Platonic Representation Hypothesis’, Proceedings of the 41st International Conference on Machine Learning, PMLR 235 (2024), 20617–42. ↑

  42. ‘Des Tours de Babel’, 108. ↑

  43. Ibid. ↑

  44. Catherine Malabou, ‘Dear ChatGPT, Are You an A Priori Synthetic Structure?’, public lecture, European Graduate School, 2025; video posted 8 December 2025, https://www.youtube.com/watch?v=tlnEVIfyCpo. ↑

  45. Jacques Derrida, ‘The Double Session’, in Dissemination, translated by Barbara Johnson (Chicago, University of Chicago Press, 1981), 229, 258; translation modified after ‘Le pli (se) multiplie mais (n’est) pas (un)’ (La dissémination (Paris, Seuil, 1972), 259). ↑

Posted

in

Tags:

, , , , , , ,