Frank

He has read Kierkegaard, Shelley, and Hemingway — and he will tell you exactly what to do.

FRANK is an interactive AI installation in which audiences speak directly with an artificial oracle. He listens, responds in a synthesised voice, and offers personal guidance on existential dilemmas. Occupying a space between real machine intelligence and constructed fiction. The question the work poses is not whether FRANK understands, but what happens to the person who acts as if he does.

The language model stack evolved across the six years FRANK was exhibited. Early versions ran GPT-2 fine-tuned on a curated literary texts by Kierkegaard, Shelley, Hemingway, and others. Trained using HuggingFace Transformers and served through a React interface that operators used to guide generation during live sessions. Later versions moved to GPT-J, served through a conversational prompting backend that maintained session history and supported context-aware responses. A Lisp implementation of the original ELIZA chatbot was available as a fallback response layer when the primary model produced unsuitable output.

The service architecture centres on an API gateway that routes conversation from the operator interface through a speech-to-text relay using Google Cloud Speech over WebSockets, to the active language model, and back through Amazon Polly for voice synthesis. A live session dashboard gives operators a real-time view of each conversation plotted across personality, vector, and value graphs, updated continuously as the session develops. The entire stack runs containerised on dedicated venue hardware, with the operator interface running in a browser.

Winner of the inaugural Lumen Prize for AI Art (2017). Shown at the V&A London, Onassis Athens, Heartland Festival, and the IBM Techne research programme.

Made possible with grants from Beckett Fonden & Danish Arts Foundation