“In a just world, every article about GPT-4 would nod toward Barthes and Foucault.”
— Ted Underwood, The Empirical Triumph of Theory
I’m late to the party, I know, but I recently became aware of Ted Underwood and his truly awesome work. I couldn’t resist reading when I saw his post The Empirical Triumph of Theory. It was as if he wrote it specifically as bait for me, a “graduate student who fell asleep in 1982 and woke up in 2022” — AI, old-school critical theory, sweeping arguments. How could I resist?
Digging in, I was taken by Ted’s cogent argument, which I’ll oversimplify here as something like this: What large language models are delivering to us these days perfectly demonstrates what all those “PoMo” theorists have been telling us about the way language and discourse work. But: Isn’t it strange that the lefty inheritors of those critical theory mantles are lining up against AI’s compelling proof in the theoretical pudding? Instead they seem to be lining up to call AI — or what we should probably call “large language models” (LLMs) — a late-capitalist shell game rather than the most awesome confirmation of what they’ve been trying to tell us all this time…
Ted had me in the first part of this argument: The LLM stochastic parrots do seem like they demonstrate how “continental” critical theory heads (like me) imagine the inner workings of discourse — as Ted describes it: “the thesis that language is not an inert medium used by individuals to express their thoughts but a system that actively determines the contours of the thinkable.” It’s as if LLMs are machines specifically designed to prove this point: For our every inquiry, they give us authoritative-sounding answers that are literally made out of their best statistical guesses at what recorded knowledge and culture have already contoured them to render as “thinkable” — which, for LLMs, means expressible, because they do not think they way you think they think.
But I took a pause in my breathless relish (?) for this initial point to ponder the deeper (?) conundrum Ted brings up around why lefty critical theory folks might not be embracing AI as a vindication of what they’ve been saying all along.
And nod toward Barthes and Foucault I do, because my newly awoken 1982-era training leads me to broaden the field of view. Yes, if I accept the AI-industry conceit that LLMs are manifesting consciousness, and limit my focus to only the output they generate, I can see a powerful analogy that does illustrate how discourse operates from a post-structuralist viewpoint. But if I step back and look at LLMs in the broader field of power/knowledge in which they have arisen, I immediately understand why a 20th-century lefty critical theorist might want to throw a spanner in their works — wait, what is the French version of that saying?
I don’t want to get all historical materialist on you (or do I?), but even the most cursory glance at LLMs has to recognize how their answers are generated via wider flows of power/knowledge beyond mere language: If you go beyond the chatbot as Socrates that AI marketing wants us to embrace, you immediately see how LLMs are constructed out of very specific collections of data, shaped by very specific models, deployed via very specific means, made possible by very specific technological and human resources, and funded by very specific interests — all of which are highly discursive (and materialist) activities. If late capitalism is pushing my hero Michel Foucault into my face with AI, I should definitely be asking why.
All that said, what I would ask is that we not take AI/LLMs at their word — if you will — and instead think about how they may demonstrate theory, not as authors Barthes already deemed dead in 1967, but as eruptions of material discourse in our own age.
I also don’t quite get the swerve in Ted’s essay from PoMo — or what I would call post-structuralist — theorists like Foucault to their structuralist antecedents like Saussure, Jakobson, Levi-Strauss, et al, although I do agree there’s maybe an investigation to be conducted that examines the relationship between structuralist ideas of discourse as metonymy and metaphor and how LLMs construct discourse statistically. But I’ll leave that to whoever thinks it would be interesting. After all, it’s not like I personally am actually going to add anything constructive to this conversation ????
Meanwhile, whatever else this essay did, it made me a diehard Ted Underwood fanboy, breathlessly awaiting his next work.