Necrotextual Poetics: Fine-Tuned Language Models and Oscar Wilde's Unfinished Aphorism
Proposal Type
Individual Talk
Location
Algorithms & Imaginaries
Start Date
July 2026
End Date
July 2026
Abstract
Oscar Wilde, in an epigraph to an unpublished manuscript, drafted the fragment of an aphorism, "To be thoroughly evil, one must..." Because Wilde never finished the sentence or documented his intention elsewhere, we cannot be certain what conditions he wished to impose nor for that matter any specific thresholds for evil. The hyperbolic qualifiers, "thoroughly" and "must" could be setting up an ironic juxtaposition, which would be typical of Wilde's style, as when he writes "There is only one thing in the world worse than being talked about, and that is not being talked about" or when Lord Henry proclaims to Dorian Gray that, "There are only two kinds of people who are really fascinating--people who know absolutely everything, and people who know absolutely nothing."
Upon discovering the unfinished aphorism, researcher Sandra Leonard posed it to a group of Wilde scholars, including my colleague Chris Foss, to see what they thought Wilde intended to write. Since predicting text completions based on latent data derived from a corpus is the core operation of large language models, this seemed like a job for generative AI. In this presentation, I will share results of an experiment in "necrotextual poetics" (a collaboration with Sandra Leonard and Chris Foss) that uses the technonecromancy method introduced in a workshop at ELO25. By fine tuning a suite of language models -- including open source models and those that restrict their training data to public domain sources -- I charge these Wildean revenants with completing his sentence.
Wilde is a uniquely interesting subject for technonecromancy both because of his interest in the macabre and because, after his death, a psychic named Hester Travers Smith published new work that she claimed had been transmitted to her via psychic messages from Oscar Wilde. Arthur Conan Doyle found these works so unmistakably Wilde-like in style that he accepted Smith's claims as evidence supporting his belief in spiritualism. Whether Gemma, Mistral, Comma, or their kin are quite as convincing will be evaluated in the second phase of the experiment, which will subject these generated and psychicically-derived works to stylometric evaluation.
While the present research is not on electronic literature per se, it is an application of computationally creative tools towards a better understanding of a literary mystery.
Necrotextual Poetics: Fine-Tuned Language Models and Oscar Wilde's Unfinished Aphorism
Algorithms & Imaginaries
Oscar Wilde, in an epigraph to an unpublished manuscript, drafted the fragment of an aphorism, "To be thoroughly evil, one must..." Because Wilde never finished the sentence or documented his intention elsewhere, we cannot be certain what conditions he wished to impose nor for that matter any specific thresholds for evil. The hyperbolic qualifiers, "thoroughly" and "must" could be setting up an ironic juxtaposition, which would be typical of Wilde's style, as when he writes "There is only one thing in the world worse than being talked about, and that is not being talked about" or when Lord Henry proclaims to Dorian Gray that, "There are only two kinds of people who are really fascinating--people who know absolutely everything, and people who know absolutely nothing."
Upon discovering the unfinished aphorism, researcher Sandra Leonard posed it to a group of Wilde scholars, including my colleague Chris Foss, to see what they thought Wilde intended to write. Since predicting text completions based on latent data derived from a corpus is the core operation of large language models, this seemed like a job for generative AI. In this presentation, I will share results of an experiment in "necrotextual poetics" (a collaboration with Sandra Leonard and Chris Foss) that uses the technonecromancy method introduced in a workshop at ELO25. By fine tuning a suite of language models -- including open source models and those that restrict their training data to public domain sources -- I charge these Wildean revenants with completing his sentence.
Wilde is a uniquely interesting subject for technonecromancy both because of his interest in the macabre and because, after his death, a psychic named Hester Travers Smith published new work that she claimed had been transmitted to her via psychic messages from Oscar Wilde. Arthur Conan Doyle found these works so unmistakably Wilde-like in style that he accepted Smith's claims as evidence supporting his belief in spiritualism. Whether Gemma, Mistral, Comma, or their kin are quite as convincing will be evaluated in the second phase of the experiment, which will subject these generated and psychicically-derived works to stylometric evaluation.
While the present research is not on electronic literature per se, it is an application of computationally creative tools towards a better understanding of a literary mystery.
Bio
Zach Whalen is an Associate Professor at the University of Mary Washington where he teaches creative coding, game studies, graphic novels, and electronic literature. He is the co-editor of Disability in Comic Books and Graphic Narratives and of Playing the Past: History and Nostalgia in Video Games. As a practitioner of computational writing, Whalen has published tiny digital poems in Taper, several popular artistic and literary Twitter bots (all of which are currently defunct), and the graphic novel An Arthrogram. He is currently working on a scholarly monograph about computer-generated books and the Electronic Literature Collection, Volume 5.