FLOR 
(Florilegium of Latent Operations and Representations)

Ongoing
The Descent. T. S. Eliot’s “The Dry Salvages” (Four Quartets) run through GPT-2 small. The original poem floats on top like a ghost and beneath it, the model’s second and third guesses at each word form new poems that exist in the model’s internals but that we usually never see.


FLOR (Florilegium of Latent Operations and Representations) stems from a concern for technical participation with a focus on large language models (LLMs) and their materiality. LLMs are increasingly used in reading practices, and often present a single type of interface, the chat box, which masks the internal operations and representations of these systems. This type of interaction tends to generate a fluency bias that reduces opportunities for interpreting and contesting their internal operations.

FLOR presents a way to engage LLMs’ formal materiality through a toolkit of fifteen lightweight software tools that use LLMs’ intermediate representations, such as token embeddings, activations, and attention weights, as material for generating poetry. The project is grounded in 20th century artistic experiments in form and language such as concrete poetry, Mallarmé’s typographic field, Oulipo, asemic writing, and the long tradition of experimental layout and book making that culminated into the established genres of artist’s book and zines.

FLOR’s tools use poems as a readout of model internal states, making visible and experiential their intermediate representations. For instance, one of FLOR’s tools, Superposition, makes visible the probability distribution of GPT-2 small by declining a source text into a series of poems based on the next-most-likely token alternatives. The result is then printed with a pen plotter whose toolpaths are modulated by the model’s quantities, whereas stroke speed and pen pressure track the probability of each sampled token. The plotter therefore inscribes the model’s certainty as a physical property of the mark.

The outputs of the project will be: 1) a series of open-source software tools that enable interactions with large language models’ intermediate representations through poetry crafting; 2) a workshop on the materiality of reading practices in the age of LLMs; 3) a peer-reviewed publication; and 4) a poetry anthology. FLOR complements conventional approaches in explainable AI (XAI), which focus on explanations, whereas FLOR emphasizes manipulation and experimentation.


Work  Home