Our session aims to describe the currently increasingly critical distinction for our humanities profession between the multiform use and integration of words (characters, images and signs of every kind) by human beings within the course of their experience and interaction with one another, and mathematically inferred determinations of the probable informational value of any word or word sequence within any equally inferentially determined context. Manually "weighted" by ai data minders to reproduce and associate digitized words and images according to the frequency of their already digitized association, or probability, "large language models" not only mistake quantity for quality (of contents, "content creators,"
and receivers) but
must be incapable of discriminating in real time between reports of empirical fact and every variety of fiction, including the fictions or "hallucinations," which, triggered by overlaps between "banked" digital sequences, must increase in probability with the accumulation of these. In short: since "llms" record bits rather than perceive phenomena, everthing that "appears" to llms appears as information. In addition, because that information is data, "givens" of the past, the probability of its selection and assocation is also determined by given (past) usage. For, as much as anti-humanists may pretend otherwise, like things, data do not "speak for themselves," have no metalanguage of their own with which to communicate the most distinctive dimension of any articulation, meaning. And while its effect upon philosophy of language was nothing short of revolutionary, no user of language required the constitution of a
theory of speech acts to use language to "act" -- that is, to perform or represent a
meaning -- in ways that dictionaries can only attempt to define after the fact. With meaning come ambiguity of meaning and grammar, metaphor, irony, allegory, symbol, "word play," and every other figural and sensory shaping of meaning by other means than the necessarily regressive rule of inferred probability. As old as Aristotle -- original advocate of the rule of
mimesis by "probability" inferred in turn on the circular basis of
endoxa or already commonly held opinions -- and as new as techno-capitalism -- extractor of nearly incalculable profit from the full range of human labor now extended to include every thought, desire, aspiration, question, and work of imagination of any dimension ever articulated by any human and recorded by any software platform anywhere -- the doubly neoclassical "logic" of llms makes for unexpected bedfellows.
For like many pretended "leaps" forward, llms define the "future" by a reductive version of the past, and the improbable -- finally, the only possible content of actual change -- as impossible by predetermined assumptions. They do this at a rate that makes of speed itself the ultimate and ultimately mystifying commodity, "magical" reincarnation of the god of medieval mystics. By ultimately equating the "products" of all human thought with physical information in the mechanical sense, llms are good for business -- for those reaping billions from the parasitic business of expropriating the entire universe of articulate human labor -- and bad for the rest of us, when information resulting from purely physical laws of causality -- data that is indeed instrumental to the advance of medical, astrophysical, geological and other forms of applied "engineering" and research -- is not the commodity we are seeking.
While the arithmetic of llms does differ both in speed and quantity from the encyclopedic models for compiling information of the past, mathematical positivism and the development of ever more effective means for its application have been around since people began learning signs by heart. The mechanistic aspects of the learning process, highlighted by such materialist "romantics" as Hegel and Wordsworth, have long been explored in literary texts. For, literature, the site of imaginative articulation, must take seriously not only the consequences of negating its reality by conflating fiction with data, but those of every confusion of machine "learning" -- reproduction and "parroting" -- with actual learning, which must instead likewise involve imaginative acts of abstraction and interpretation. As experts in artificial so-called "intelligence" freely admit, llms are by definition incapable of judgment and interpretation, intellectual action rather than mimicry.
Reflecting these critical concerns, panelists Chandra and Sanazaro analyze exposés of machine-learning across diverse imaginative works extending from Ghalib to Dickens, Beckett, and Derrida, while Brodsky distinguishes how words work from data in proactive defense of the acts of reading, interpreting and teaching imaginative literature. While we rightly debate how to do these, it is the doing itself that is now at issue.
