p. 48 on iterability (important to understanding constraints of large language models)

« After all, there is little doubt that human languages can realize logical forms, or that some parts of linguistic practice appear logical on the sur. face: this is the observation that licenses almost all linguistic approaches that are interested in form. But human language can be used to work mathematical and overtly logical systems like the formal logic taught in contemporary philosophy classes, and even programming languages them. selves: the fact that language is capable of simulating these systems cannot be taken as strong evidence that language is such a system. The parts of language that escape formalization are well understood and have long been recognized as problems for pure and purely autonomous theories of form; these are exactly the phenomena that Derrida calls iterability, and the presence of idioms. Iterability in this narrow sense points to the fact that any linguistic object can apparently be repurposed for uses that diverge from what appear to be the syntactic compositional elements of a given ob-ject. Thus while the phrase "snow is white," so often invoked by logicians, appears to have a single, stable worldly denotation, in practice it can be used to mean an almost infinite number of other things, and in fact such a bald declarative statement would rarely, in human language practice, be used simply to express its apparent denotative meaning. Even obvious and repeated phrases like "hello" and "goodbye," via processes of iteration and citation, often bear much more meaning than they would seem to do from the syntactic view, and can be iterated for other purposes, if we can even say what "the meaning" of words and phrases are in any given context. »