This project is funded by the Alfred P. Sloan Foundation
in partnership with the United Kingdom Metascience Unit and Canada’s Social Sciences and Humanities Research Council
Interpretative Interfaces
Design Research
For centuries, the book was a difficult technology. Producing a single manuscript required months of skilled labor, such as scribes painstakingly copying letter by letter onto prepared animal skins, grounding and mixing pigments by hand, and stitching and pressing bindings. The cost of a book could rival that of a house or a plot of land. Even after the printing press was invented, the price of paper, the weight of metal type, and the economics of distribution kept books out of most people's hands for generations. Literacy was a technical skill restricted by class, wealth, gender, and access to instruction.
What made books progressively more legible to wider audiences wasn't only the printing press or smoother paper, but the slow development of margins, interlinear gaps, glosses, and marginalia––that is, the development of spaces where readers could intervene, and comment, discuss with, and possibly contest the text.
Today's large language models are in their pre-margin era. Chat windows and explanation dashboards tell users what a model did; they rarely let users work with how it did it. And while explanation are complementary, what is needed are opportunities to interpret the material of AI.
Interpretative Interfaces is an exploration of what it would look like to make engagement with an LLM's intermediate layers as fundamental to the interaction as prompting. We present a series of prototype interface designs that treat these intermediate layers as a space for engagement. The prototypes are objects to think with; proposals for what AI interfaces could become if we designed them for interpretative engagement.
What made books progressively more legible to wider audiences wasn't only the printing press or smoother paper, but the slow development of margins, interlinear gaps, glosses, and marginalia––that is, the development of spaces where readers could intervene, and comment, discuss with, and possibly contest the text.
Today's large language models are in their pre-margin era. Chat windows and explanation dashboards tell users what a model did; they rarely let users work with how it did it. And while explanation are complementary, what is needed are opportunities to interpret the material of AI.
Interpretative Interfaces is an exploration of what it would look like to make engagement with an LLM's intermediate layers as fundamental to the interaction as prompting. We present a series of prototype interface designs that treat these intermediate layers as a space for engagement. The prototypes are objects to think with; proposals for what AI interfaces could become if we designed them for interpretative engagement.
Left: Hidden-State Projection, 2026. A map of E. E. Cummings’s “in Just-” inside GPT-2. All 768 dimensions flattened into two. Every dot represents a word and the lines show how the words move, layer by layer. Right: Halftone Plate, 2026. The same points as a halftone plate. Dark areas show where a word’s representation changes most from layer to layer, that is, where the model “rewrites” most.
Entropy Knits, 2026. A series of tangible interfaces. Left: Per-token entropy for “Roses are red, violets are blue” as a knitted piece. The darker areas show where the model is less sure. Right: Per-token entropy for “All Watched Over by Machines of Loving Grace.” This line produced many levels of uncertainty for GPT-2, and each mapped to a different needle change. The frequent changes caused the knitting machine to jam.