Neo-structuralism

This recent New York Times article about Netflix looks at the persistence of tagging as a profitable way for Netflix to organize its content, offering users the most expedient way to achieve a sense of algorithmic fulfillment.

The two- or three-word tags, meant to convey the gestalt of a show or movie, regularly help viewers choose a show from the service’s nearly endless library, the company says. The words are selected by about 30 employees — so-called taggers.

(That reminds me of when we used to tag Real Life stories; it was an elaborate and extremely time-consuming process.)

The tags are necessary because, as the article flatly asserts, “picking a show or movie is often tedious and frustrating.” Rather than choose among the things themselves — the “more than 10,000 titles on Netflix and thousands more on other streaming services” — tags let users choose among a smaller subset of affects, vibes, microgenres, and abstract concepts. Shopping for feelings.

I find this sort of thing strange, because I usually want to watch something specific or else I won’t turn the TV on. The article quotes a marketer who claims that “when we see the term ‘gritty’ or we see the term ‘cerebral,’ we understand intrinsically what that means,” but I don’t know if I believe that. What do people want out of passive viewing? Do those tags describe the content or are they meant to capture what we think of ourselves? Do we watch to learn what grittiness or cerebrality looks or feels like, to calibrate our conceptual understanding? Is there a connection between the system of tags and the system of identity? This assertion of an “intrinsic understanding” seems like an attempted disavowal of how the terms don’t really refer to anything specific, but are just differential signifiers in a relational system sustained by Netflix within Netflix.

As my last few posts probably indicate, I’ve been reading up a bit on structuralism, trying to see whether the controversies about it and the various critiques of its aims and methods offer anything for thinking about LLMs. It seems to me that language models, maybe by definition, are automated versions of structuralism, in that they construct a synchronic, weighted matrix of language’s possibilities that is forcibly detached from history — a model of la langue that can generate an infinite number of paroles. And some structuralists, like some “AI” advocates today, were eager to marginalize human consciousness as a unique constitutive force — thus you find Claude Lévi-Strauss declaring in La Pensée sauvage that “the final goal of the human sciences is not to constitute man but to dissolve him.” This entails “reintegrating culture into nature and finally life into the set of its physico-chemical conditions.” This seems in keeping with AI’s aim of predicting the future on the basis of past data — once you sufficiently understand the physico-chemical conditions of the world, you should be able to program it (and all the living beings it comprises) like a computer.

In Fredric Jameson’s “Metacommentary” (which serendipitously popped up on one of my social media feeds last week), he describes structuralism as

formalistic in that it studies organization rather than content, and assumes the primacy of the linguistic model, the predominance of language and of linguistic structures in the shaping of meaningful experiences. All the layers or levels of social life are ordered or systematic only insofar as they form languages of their own, in strictest analogy to the purely linguistic: styles of clothing, economic relationships, table manners and national cuisines, kinship systems, the publicity apparatus of the capitalist countries, the cosmological legends of primitive tribes, even the mechanisms of the Freudian mental topology — all are systems of signs, based on differential perceptions, and governed by categories of exchange and transformation.

If structuralism assumes everything is structured like a language, that would suit the tech company ideology of developing LLMs and presenting them as the path to “artificial general intelligence.” What Jameson discusses as “linguistic structures” we might think of now as datafication or digitization, processes predicated on the idea that there is a systematic order to “all the layers of social life” that can be captured quantitatively, and that all experience is essentially fungible, exchangeable, commensurate.

But the “systematic order” of language also appears to be a bit random. Ferdinand de Saussure famously claimed that the relation between signifier and signified was arbitrary, that language derived meaning from the opposition of otherwise meaningless tokens. There is no necessary pattern for matching sets of sounds with particular things in the world, yet people have collectively assigned meanings to the distinctions between those sets of sounds. In Jameson’s terms, structuralism is “anti-substantialist.” It locates meaning in the difference between contrasting elements and not the thing in itself, so “substance is replaced by relationship,” and “all meanings are organized, following the pattern of phonology, in pairs of oppositions or determinate differences.”

The question structuralism raises is to what degree we can interpret why certain meanings are assigned to certain determinate differences — why certain words are assigned to certain things, or more broadly, what links the more or less coherent system of language (what the LLM seeks to approximate) and the systems that order things in the world. Are these homologous? Do the relations mapped in language have bearing on the relations outside of language, and if so, how?

In the Introduction to the Work of Marcel Mauss, Lévi-Strauss seems to say the essence of human knowledge is to progressively undo the arbitrariness of the relations between signifier and signified.

the two categories of the signifier and the signified came to be constituted simultaneously and interdependently, as complementary units; whereas knowledge, that is, the intellectual process which enables us to identify certain aspects of the signifier and certain aspects of the signified, one by reference to the other — we could even say the process which enables us to choose, from the entirety of the signifier and from the entirety of the signified, those parts which present the most satisfying relations of mutual agreement — only got started very slowly. It is as if humankind had suddenly acquired an immense domain and the detailed plan of that domain, along with a notion of the reciprocal relationship of domain and plan; but had spent millennia learning which specific symbols of the plan represented the different aspects of the domain.

I find that description very strange: It is as if he is claiming that we were given “language” and the “world” at the same time in their totality, and if we figure out why they are related, we will have solved the riddles of our existence. But the it fits with the maximalist aspirations for LLMs: that the statistical patterns of reference revealed in language use en masse aren’t merely conventional or historical and subject to constant revision, but instead they reveal some basic a priori truth about the human condition, the universe as it has been given once and for all and how it really works.

It’s reminiscent of the moment in Kant’s Critique of Pure Reason when he claims that there are “no rules for judgment” — i.e. no rules for what concepts apply to which things — and that “it is a peculiar talent which can be practiced only, and cannot be taught. It is the specific quality of so-called mother-wit; and it’s lack no school can make good.” No class could teach you how to be a good Netflix tagger.

What allows us to know a dog when we see one? Kant says:

 The concept of a dog signifies a rule in accordance with which my imagination can specify the shape of a four-footed animal in general, without being restricted to any single particular shape that experience offers me or any possible image that I can exhibit in concreto. This schematism of our understanding with regard to appearances and their mere form is a hidden art in the depths of the human soul, whose true operations we can divine from nature and lay unveiled before our eyes only with difficulty.

Structuralism seems to want to reveal this “hidden art in the depths of the human soul” and “divine from nature” how and why certain things and concepts are associated. Lévi-Strauss claims that

the universe signified long before people began to know what it signified; no doubt that goes without saying. But, from the foregoing analysis, it also emerges that from the beginning, the universe signified the totality of what humankind can expect to know about it. What people call the progress of the human mind and, in any case, the progress of scientific knowledge, could only have been and can only ever be constituted out of processes of correcting and recutting of patterns, regrouping, defining relationships of belonging and discovering new resources, inside a totality which is closed and complementary to itself.

Viewing the universe as a closed totality suggests that time itself is meaningless, an unfolding of a pre-existing and unalterable pattern that humans try to fathom but can’t change. It seems as though this vision animates the claims and the fears of the companies working to impose language models on us as the apotheosis of general intelligence.