Literary Cybernetics: The Point (of the Spear)
N. Katherine Hayles · New Literary History · 2023
Literary Cybernetics:The Point (of the Spear) N. Katherine Hayles (bio) Cybernetics and literary studies are on a collision course that will transform what it means to read, to write, and to be human. The essays on "literary cybernetics" in this issue touch on many different ways in which this phrase can be interpreted, but for all their rich variety, they do not entirely capture either the urgency of our present situation or the inevitability of the coming transformations, even if we do not know and cannot reliably predict exactly what new forms will emerge, postcollision. At the pointy end of the spear driving into the heart of literary studies are the large language models (LLMs) created by rich tech companies and increasingly available to the general public such as OpenAI's GPT-3, -4, and ChatGPT (Generative Pretrained Transformer, versions -3 and -4), Google's LaMDA (Language Model for Dialogue Applications), and Google's BERT (Bidirectional Encoder Representations from Transformers). GPT-3, for example, was trained on forty-five terabytes of human-authored texts, mostly scraped from the web (a terabyte of data would fill about 570 million pages).1 Reading for GPT-4 is very different than for a human child learning to decode letters. Words are broken into tokens (generally word fragments of about four letters) and transformed into vectors processed through its ninety-six layers of neurons. This generates probability matrixes which are then processed further through a software function such as Softmax and output as words. The outputs are probabilistic projections of what the next word (or series of words) in a sequence would be.2 These models instantiate many of the cybernetic concepts discussed in this volume's essays, including the qualities emphasized in Paul Jaussen's "The Art of Distinction." They include recursivity (outputs are fed back into the model as inputs) and environment/system distinctions. Originally the training set of texts constitutes an LLM's environment, but as the model learns, assumptions in the data set are absorbed into the continuously adjusted weights of the different neuron layers; after training, the model's interactions with human interlocutors constitute another kind of environment/system distinction. So intricate are the model's instantiations of these and other [End Page 1289] cybernetic ideas that the models can legitimately be called the ultimate cybernetic machines. The textual outputs of GPT-4 are several orders of magnitude more sophisticated than those produced by chatbots such as Siri and earlier algorithmic text-programs. GPT-4's texts are not only syntactically correct and semantically coherent; they often also demonstrate dazzlingly complex turns of rhetoric and logic. Indeed, they are often so good that they cannot be reliably distinguished from human-authored texts. From the technical description above, it would be quite surprising to discover that GPT-4 can discern and reproduce literary styles as diverse as the King James Bible and Mark Twain, and even more surprising to find that it can also identify and reproduce high-level literary qualities such as genre. Essays written by GPT-3 have been submitted to college professors for evaluation as if they were written by undergraduate students and have received excellent grades (A and A-), along with positive comments.3 As Lea Pao notes in "Ways of Cybernetic Thinking," one problem that immediately leaps to mind for educators is the chaos this will cause for humanities courses, where the typical way in which students demonstrate their abilities to synthesize and interpret what they have learned is by having them write essays on assigned topics. As disruptive as this looming deluge of plagiarism may be, in a philosophical sense it is the least of the transformative effects. More fundamental are debates over what kinds of meaning machine-produced texts can have, and indeed if they can be said to have any meaning at all, beyond what human readers project into them. For literary studies, the blurring of boundaries between human cognition and machine learning means that assumptions about the uniqueness of the human ability to learn, use, and manipulate symbolic abstractions and languages are immediately drawn into question.4 Already machine-generated texts are ubiquitous in formulaic prose such as sports...