AIs Need Time To Work - Derrida and LLMs

Bashir argues for structuralism's re-emergence, I argue there's no time.

Daniel Bashir’s The Third Yes essay is a thoughtful analysis of current AI, but it is a recovery project, one that dances close to irredentism by using the transformer architecture to reclaim structuralist territory. Post-structuralism becomes, for Bashir, a cultural cul-de-sac—a picturesque distraction. The appeal makes sense. Language models make structure great again. But Structuralism was superseded because it could not account for time, responsibility, and use, and those pressures haven’t vanished just because we can train a transformer. Post-structuralism took them up, and they remain the fitting program for the next phase of generative AI.

Treating an LLM’s frozen state as revealing something fundamental about language is a category error. It is an engineering limitation, not ontology. Current LLMs are massive weightings derived from static word-token maps; they reflect Saussure’s langue because they are fixed systems of differences, but parole never rewrites the system. Forward time—essential to language—remains unaccounted for.

The unreasonable language facility of LLMs makes a structuralist revival feel justified, but that risks making a feature of a flaw. The model, even when supple within its context window, is composed of calcified weights. The next phase—the “unheard-of computer”—will require exactly the kind of dynamic reweighting current LLMs lack. Until then, these systems are constellation maps of a night sky in a frozen universe.

Real language, as the Post structuralists intuited, changes over time. LLMs don’t—at least not through conversation itself. We ignore this because they can represent past change so convincingly, producing fake Elizabethan sonnets and simulating long-concluded language games.

Bashir’s “third yes” thus risks becoming a clever gramophone: a sophisticated See and Say or Teddy Ruxpin that replays old linguistic games with simulated drama but nothing at stake. When generation is layered atop a static model, the words change, but the weights do not. Nothing in the system risks modification.

The “Platonic” convergence Bashir describes is fascinating but likely inevitable in any closed-loop system. Current transformers optimize frozen parameters against static data. What comes next—active inference or something like it—will have to adjust continuously, staying in touch with a world that surprises it. A model can only countersign, to return to Derrida, if it can be altered by the encounter. GPTs remain gramophones—technically incapable of the processual transformation that defines real language—until we move from the static architecture of the transformer to a dynamic system where learning and inference are the same continuous movement.

One of the most satisfying aspects of the transformer revolution is that it has revived debates once dismissed as desiccated—the debates that produced both Structuralism and Post-structuralism. What many experienced as frustrating academic wordplay in the twentieth century has become a living space of inquiry. But midwifing the offspring of critical theory and late cybernetics requires deliberateness, and dismissing a branch of either discipline should be done with care.

We need to stop focusing on the cave and start understanding what it would mean to walk out of it.