Perhaps something has occurred in the history of the concept of structure

In “Structure, Sign, and Play in the Discourse of the Human Sciences,” the 1966 lecture often credited with launching post-structuralism, Jacques Derrida quotes a passage about the advent of language from Claude Lévi-Strauss’s Introduction to the Work of Marcel Mauss:

Whatever may have been the moment and the circumstances of its appearance in the scale of animal life, language could only have been born in one fell swoop. Things could not have set about signifying progressively. Following a transformation the study of which is not the concern of the social sciences, but rather of biology and psychology, a crossing over came about from a stage where nothing had a meaning to another where everything possessed it.

It’s tempting to construe the training of LLMs in these terms: Open AI’s huge pile of stolen data is just an inert mass of nothing until the model is switched on and it “crosses over.” Suddenly the pile of nothing is fully structured and endlessly generative, a machine for enabling “play.” We don’t have to speculate about how it happened or why; instead we are invited to project the programming of neural nets back onto human “biology and psychology” and consider Lévi-Strauss’s problem solved.

In La Pensée sauvage, Lévi-Strauss even “evoked a program reserved for the ethnology of a future century” — “a classification of classifications” that “could be realized only on the condition of searching so many documents and taking account of such varied dimensions that even in restricting oneself to societies for which data are sufficiently rich, precise, and comparable among themselves, one could not do it without the aid of machines.” It seems as though those machines are here.

Unlike with language itself, we don’t have be nostalgic over the lost origin point of LLMs; we don’t have to think of them as problematically ahistorical. We know precisely when and how they cross over, even if we can’t fully explain the logic behind their play. In that sense they could be seen as post-post-structural, or contra-post-structural: They could be taken to address or to obviate some of the concerns Derrida has about structuralism and authorize us to proceed as though structuralism is no longer a problematic methodology.

Part of Derrida’s critique of Lévi-Strauss is that his sort of structuralism requires bracketing off history and indulging a nostalgia for the lost origin point, which is also the missing center that organizes the otherwise arbitrary set of relations in language together. “By orienting and organizing the coherence of the system, the center of a structure permits the play of its elements inside the total form. And even today the notion of a structure lacking any center represents the unthinkable itself,” Derrida writes. The idea of the structure’s center secures “the concept of a play based on a fundamental ground, a play constituted on the basis of a fundamental immobility and a reassuring certitude, which itself is beyond the reach of play.”

But that center, Derrida argues, is always absent, always paradoxical, “contradictorily coherent”: it is part of the structure but also outside it. When Lévi-Strauss and the like begin to challenge Western ethnocentrism, the subject’s self-understanding, and Hegelian myths of historical progress, and think “the structurality of structure,” as Derrida puts it, the decentered center could no longer be overlooked.

Henceforth, it became necessary to think both the law which somehow governed the desire for a center in the constitution of structure, and the process of signification which orders the displacements and substitutions for this law of central presence—but a central presence which has never been itself, has always already been exiled from itself into its own substitute. The substitute does not substitute itself for anything which has somehow existed before it. Henceforth, it was necessary to begin thinking that there was no center, that the center could not be thought in the form of a present-being, that the center had no natural site, that it was not a fixed locus but a function, a sort of nonlocus in which an infinite number of sign-substitutions came into play. This was the moment when language invaded the universal problematic, the moment when, in the absence of a center or origin, everything became discourse—provided we can agree on this word—that is to say, a system in which the central signified, the original or transcendental signified, is never absolutely present outside a system of differences. The absence of the transcendental signified extends the domain and the play of signification infinitely.

In this recent post by Henry Farrell, about LLMs as “cultural technologies,” he references Derrida’s maxim that “there is no outside the text,” which is one of the implications of this. Nothing outside language (a.k.a. “the real world”) ultimately grounds or guarantees the relations revealed or articulated through language, which are generated by the playing out of the system itself. One of the striking things that Lévi-Strauss discusses in La Pensée sauvage is how empirical differences observed in the natural world are adopted as arbitrary signifiers in various classification systems; they speak to difference in the abstract. Derrida quotes his stated intention, in The Raw and the Cooked, “to transcend the opposition between the sensible and the intelligible by operating from the outset at the level of signs.” But Derrida’s point is the effort to transcend the opposition also reinscribes it.

LLMs proceed on a similar logic to that in the Lévi-Strauss quote, that one can bracket away the issue of how the system of language and the relations and oppositions to which it gives free, generative play actually link up to the world of things and to what we sense and experience through our bodies. That is how he can say language simply emerges all at once without having to trace the history or logic or contested stakes of that emergence. LLMs invite users to make a similar move, to accept a model’s analysis of the structure of language (as it exists in a more or less arbitrary collection of samples) independent of an analysis of how and why that structure took shape in its particular form, with particular sociopolitical implications.

The companies purveying LLMs hope to forestall how, in Derrida’s words, “language bears within itself the necessity of its own critique.” The language model is put forward as though it has solved these conceptual conundrums of language, presenting language in a final, solved state. Lévi-Strauss averred that “in no instance would I feel constrained to accept the arbitrary demand for a total mythological pattern, since, as has been shown, such a requirement has no meaning.” You can’t exhaust the possibilities of la langue, so coming up with a model on the basis of the data you managed to have around was enough. The makers of LLMs have taken this to heart.

Yet at the same time, Lévi-Strauss has to acknowledge the ad hoc nature of structural analysis. Derrida quotes a passage where he admits that with the study of myth,

there is no real end to methodological analysis, no hidden unity to be grasped once the breaking-down process has been completed. Themes can be split up ad infinitum. Just when you think you have disentangled and separated them, you realize that they are knitting together again in response to the operation of unexpected affinities. Consequently the unity of the myth is never more than tendential and projective and cannot reflect a state or a particular moment of the myth. It is a phenomenon of the imagination, resulting from the attempt at interpretation; and its function is to endow the myth with synthetic form and to prevent its disintegration into a confusion of opposites.

He sounds like someone trying to explain an LLM’s hallucinations. But tech companies are effectively insisting that they don’t need explanations, that they are beside the point. They proceed as though “big data” and the “end of theory” are enough to justify the validity of the output. They can do brute-force structuralism without reservations.

But structuralism’s status as a kind of science is still nonetheless problematic. They aspire to be empirically grounded even as the concept of structuralism itself does away with empiricism. (Derrida: “The structural schemata are always proposed as hypotheses resulting from a finite quantity of information and which are subjected to the proof of experience.”) Models organize information without the possibility of recourse to empirical observation or confirmation. Farrell cites this paper by Eunice Yiu, Eliza Kosoy, and Alison Gopnik to explain that LLMs are essentially locked in the prison house of language.

LLMs operate in a space of information that is disconnected from base reality. An LLM, doesn’t ‘know’ (to the extent that it ‘knows’ anything) that the phrase ‘soft drink stand’ refers to something that exists in the physical universe. For it, soft, drink, and stand is a series of “tokens,” individual strings of letters that don’t refer to, or have relationships with anything except other tokenized strings of letters. What the LLM ‘knows’ is the statistical weights associated with each of these tokens, which summarize its relationship with other tokens, rather than the world we live in.

LLMs don’t have embodied knowledge of anything; they are just a different form of textual archive. They only know and manipulate signs, and the relation of the signifiers to signified in these signs is arbitrary — their relations form an internally coherent system (that has no center) but that coherence has no definite relation to how things in the world are related, some aspects of which elude encoding. As Farrell notes, summarizing the Yiu, Kosoy, Gopnik paper, models “can’t ever discover these relationships in the ways that human beings can — by trying things out in the real world, and seeing what works.”

But the models can do in a new way what the mythic structures and classificatory systems Lévi-Strauss traced have always done: “cultural transmission.” “We ought pay attention to how LLMs are likely to transmit, recombine and re-organize this cultural information, and what consequences this will have for human society,” Farrell writes, which sounds a bit like doing a structural anthropology on (and/or a deconstruction of) machines’ opaque classifying processes, as well as attending to how LLMs are assimilated into existing human institutions and relations. Farrell notes that scientist James Evans has proposed using LLMs “to discover the holes in n-dimensional science space, showing us the places where there really ought be connections between researchers, and there are not,” which sounds a bit like setting Chat-GPT to work on constructing the Greimas squares to end all Greimas squares.

At the end of his lecture, Derrida names “two interpretations of interpretation, of structure, of sign, of play.” What Farrell and the academics he cites are talking about seems to correspond to the first of these, an approach that “seeks to decipher, dreams of deciphering a truth or an origin which escapes play and the order of the sign, and which lives the necessity of interpretation as an exile.” It seeks the “reassuring foundation” and sees in ethnographic practices and other methods of social science the possibility of a revitalized humanism.

The tech companies making LLMs sometimes seem to be voicing the other interpretation of interpretation, which Derrida associates with Nietzsche. It offers a “joyous affirmation of the play of the world and of the innocence of becoming, the affirmation of a world of signs without fault, without truth, and without origin which is offered to an active interpretation.” From this point of view the models aren’t machines for cultural transmission but machines of nihilism, giving users a hands-on feel for how insignificant the connections between language and reality, or nature and culture, really are to making cool shit.