"Gnarly Posthuman Conversations: John Ashbery, W.H. Auden, Wallace Stevens, and GPT-2" is a generative poetry project based on the personal and critical discourse relationship between the three poets. The project explores the potential of a "hypercarving" generative text pipeline using the relationships between the three authors as an inspiration behind the pipeline design. The pipeline starts with training two GPT-2 models, one on the poetry of John Ashbery and the other on the poetry of W.H. Auden. These two separate models are then put into "conversation" with each other, i.e. the output of one model is used as the input to the next and vice-versa in a feedback loop. The large amount of generated text is then classified using a recurrent neural network trained on each of the poets corpora. This allows for filtering the conversations in unique ways, for instance filtering the conversations by highest RNN score.
The project is presented in two ways. The first way displays the generated text in the conversation format. In this form the full conversation between the John Ashbery and W.H. Auden models is shown, along with three other filtered conversations that only display generated texts where a given author's RNN score was highest.
The second way is through a "transformative reading interface," which I think of as a concrete instantiation of Katherine Hayles' idea of a technotext. The transformative reading interface links up the original training texts of the poem of John Ashbery and W. H. Auden with the generated text based on keywords.
This piece was exhibited in the Electronic Literature Organization 2021 Festival, Posthuman E-Lit series in the library at the University of Bergen, Norway in March of 2021.