Digital theory will benefit from a return to structuralism. As I've said before in print, structuralism was the apex of digital theory during the 20th C. During that period, scholars were obsessed with codes, logics, binarisms, linguistic structures, and even quasi-mathematical approaches, as seen in the work of Claude Lévi-Strauss, or Lacan's Seminar 2 on cybernetics and binary codes. Scholars like Lydia Liu and Bernard Geoghegan have traced some of these lineages; Leif Weatherby's new book, Language Machines, takes up the question of structuralism as it relates to AI. I value their work and we ought to learn from them.
Meanwhile, problems linger, problems that attend any "return" to structuralism. I will pass over all the well-known shortcomings, already chronicled in various ways over the years: the risks of over-investing in the synchronic (at the expense of the diachronic); the rigidity of an apparatus frozen in amber (as opposed to a more dynamic or internally contradictory model); the tendency for structure to reify binary relations; structuralism's apparent Eurocentrism. Those problems remain. I'd like to address some additional shortcomings, though, ones that spring directly from digital theory.
The digital shortcomings have to do with a number of things that people care about, including meaning, value, truth, and quality. Within computation these tend to be classified as "hard problems," which is to say they don't have clear answers, and, worse, there aren't even plausible pathways for pursuing them. Then, beyond these everyday things that people care about, a number of other hard problems also remain: the transcendental (in Kant's sense), the symbolic order and the unconscious (from psychoanalysis), plus any number of other things that fall under the heading of metaphysics. It's just not clear that digital theory can account for any of these problems, much less all of them.
Let's not forget that structuralism would never have flourished without help from a number of other fields. Phenomenology, for one, but also studies of culture and practice, and doubtless the poststructuralist "supplement" that began in the late 1960s--these all saved structuralism from itself, mainly by grounding theory in a subject; but also by re-introducing all sort of para-rational and non-rational elements. Structuralism was at its best when it evolved past its own limitations (the Baudrillard of Seduction and after) or when it was caught in the narcotic spell of phenomenology (the Barthes of Camera Lucida).
If digital theory will return to structuralism it will, alas, return to a limited subset of structuralist discourse. The phenomenological or poststructuralist admixtures will be absent. Only the clinical core will remain. In this way, I see digital theory as a kind of "stunted" or "lobotomized" structuralism. The digital is a structuralism without a subject, defined strictly within a mechanism or apparatus. This is a structuralism without meaning, without the transcendental, without value, without the concept, without the symbolic order. Beatrice Fazi has a great term for this. She calls it the "subsymbolic order." Indeed, the digital doesn't have access to what psychoanalysis calls the symbolic order, except through a kind of blank, mechanical pantomime. The digital resides instead within the subsymbolic order.
Consider the question of meaning or linguistic value, one of the aforementioned "hard problems." The Saussurian tradition of structuralism defined linguistic value via differences between circulating signs. Jean-Joseph Goux, one of the great synthetic thinkers during that period, put it the following way: "Metaphors, symptoms, signs, representations: it is always through replacement that values are created" (Symbolic Economies, 9). Written signs flow, thus, as a kind of money, a currency carrying linguistic value. Jeffrey Kirkwood's essay on "Linguistic Arbitrage" is instructive here. He argues that arbitrage (exploiting differentials in value; buying low and selling high, essentially), a term borrowed from finance capitalism, is equally applicable to LLMs, specifically the value differentials, captured as vectors, that circulate within them. (I'm also reminded of McKenzie Wark's foundational discussion of vectors and the "vectoralist class.")
Meanwhile Goux, et al., have a glaring problem, the labor theory of value. Arbitrage is what Marx called mercantile capitalism, and he definitively rejected trade/mercantilism/arbitrage as a source of value. And while mainstream economic theory tries to source value from things like "marginal utility," it's capitalism itself that has confirmed the truth of the labor theory of value time and time again. The very AI companies seeking to remove humans from the loop in pursuit of "frictionless production" are themselves entirely reliant on human-generated training data. Ironically, Sam Altman is one of history's greatest adherents to the labor theory of value, just as all capitalists have been, going back to Marx's time, if not before. If capital could eliminate labor as the source of value, believe me, it would have done it by now.
So it's not that LLMs are right due to some sort of arbitrage that generates and exploits symbolic differences, but that structuralism was wrong about value appearing strictly via circulation. That's arbitrage theory; it's just not empirically valid.
In crass terms: if structuralism was correct, then LLMs are already AGI, case closed full stop. If structuralism was correct, then LLMs would generate meaning, they would harbor an unconscious, they would participate in the symbolic order. Yet, since LLMs don't do any of those things--sadly there's no ghost in the machine--structuralism has a lot of explaining to do. Yuk Hui's new book is great on this point: at best, LLMs are on par with David Hume circa 1740--at best! Evolving AI forward to the level of Kant's "transcendental unity of apperception" will be extremely difficult; getting to Hegel's dialectic will be constitutionally impossible, at least using current hardware and software.
In other words, this "return" to structuralism should be a genealogy not an archeology. I want to reach back to paleo-structuralism, not rekindle a kind of neo-structuralism for the 21st C. The legacy runs backward rather than forward. It's not just that structuralism is necessary to understand the powers of LLMs (following Weatherby). But, also, the weakness of LLMs reflects poorly, when looking in retrospect, on structuralism's overweening ambition sixty or seventy years ago. If LLMs hadn't been invented, we might have been able to slumber on blissfully, content with the science of Saussure or Jakobson. But given the existence of LLMs, the limitations of structuralism appear in a new light. Paleo-structuralism, thus, is not an exhortation toward the discoveries of the future, but an anchor dredging up the shortcomings of the past.