About Ultimate Fantasy
Ultimate Fantasy is a collection of of short stories, poetry and art generated by Queer AI, an artificial intelligence trained on a corpus of queer theatre. Every story begins with the phrase, “Let me begin by telling you my ultimate fantasy. . .”
The Collection
The collection consists of 111 unique NFTs. There is also a limited edition softcover book and audio cassette of stories from the collection.
The Artist
Emily Martinez is a new media artist, 1st generation immigrant/refugee (Cuba > Miami), and a self-taught coder who believes in the tactical misuse of technology. Their art has been exhibited internationally, mostly through collaborations with Anxious to Make and Queer AI. Their solo work is a departure from previous collaborations and a return to unburying histories of immigrant, queer un/belonging. They started the Ultimate Fantasy series while documenting some of their research and experiments with GPT-2.
TLDR; a brief history
Queer AI began as a collaboration between Emily Martinez and Ben Lerchin in 2017. They wanted to create a chatbot that embodied the messiness of relationships, bodies, and identity. The first iteration of Queer AI (v1) was trained on 50,000+ conversational pairs derived from scripts in queer theatre, using an RNN seq-to-seq algorithm. Ben wrote all the code and prepped all the data. Emily wrote manifestos and talked to bots.
The results of Queer AI were surprising, revealing much more than we had anticipated, such as the bot “inheriting” trauma from the dataset, much of which contends with the experiences of living in a queer body during the AIDs crisis.
Since then, Emily has been obsessively developing tactics for understanding a corpus through these unintended ruptures and writing stories with machine learning models that can serve up unruly, erotic, culturally-specific, counter-narratives to the erasures and bias reproduced by popular large language models.
Ultimate Fantasy started as a series of blog posts documenting some of Emily's research and experiments with machine learning during their ML5.js fellowship in 2020. At the time, Emily was also learning how to prep training data and work with machine learning models such as GPT-2, a language model created by OpenAI. A goal of their research was to study some of the bias inherent to GPT-2, which they decided to explore by finetuning the model with small datasets and then writing strategic prompts to generate texts and images that revealed some of the latent qualities and "failures" of these very large models. One of the custom models they made is an updated version of QueerAI (v2), which was used to generate all of the short stories and poems in the Ultimate Fantasy collection.
Acknowledgements
None of this would exist without amazing people inspiring, teaching, supporting and encouraging me to continue this work.