← Studies · interactive prototype · 2026

§ 3.09

Font Attention — typography as model attention

A paper-like Qwen3 attention replay mixing three typewriter faces token by token, with persistent print memory, red ballpoint emphasis, and hand-drawn contextual annotations.

Methods
Local LLM · Attention · Typewriter typography · Web interaction

The replay uses pre-recorded Qwen3-0.6B attention tensors, averaged across every head in the final four layers. Playback, scrubbing, speed control, and typeface switching all run locally in the browser without an inference server. Focus, entropy, and self-attention are mapped to typographic weight, spacing, slant, motion, and annotation as an expressive interface—not as a claim about model cognition. The default typewriter case mixes Courier Prime, Special Elite, and Cutive Mono token by token, while a selector can isolate each face. Attention links use segmented, pressure-varying ink strokes rather than Bézier curves; circles, underlines, strikeouts, and typographic disturbance hold long enough to create a calmer revision rhythm. Each incoming token connects directly back to its marked source words, and the same strongest links return as loose handwriting in the page margin.

Model
Qwen3-0.6B
Data
Recorded attention tensors · static JSON
Typography
Three typewriter faces · print memory
Controls
Replay · typeface · mapping