Alongside working with my own charRNN model, I’ve been experimenting with OpenAI’s large language model, GPT-3. It’s trained on a 45 TB dataset of texts scraped from the Internet and digitised books. I’m trying to find out how to interface with the API so that I can implement it within the surfacecollider site. Here are some of GPT-3’s offerings based on prompts written by myself, and built into a locally hosted webpage I’ve called ‘dead text’.

I trained a charRNN Machine Learning (ML) model on a dataset of texts from the bibliography of my research. The text that it generates makes little sense – there isn’t a linear unfolding of narrative structure, for example, or a development of an argument over time. It’s poetic through. Purely visual.

I’m experimenting with working the model into the surfacecollider website now – using a comment box placed in the bottom-right corner of the window to let users ‘prompt’ the model to keep writing…