Can a Computer Learn 2,300 Years of Visual Poetry?
Visual poetry has challenged readers for more than two millennia, and a remarkable new project called TYPESTRACT demonstrates how computers can preserve, rather than replace, that literary tradition.
A browser window opens and begins arranging scattered letters, shaped lines, diagrams, and worn typewriter fields into poems, and the first impression suggests another generative design toy. Closer reading reveals something more deliberate. Beneath the shifting surface sits a living archive whose contents stay active, variable, and able to produce new work on every keypress. What looks like novelty turns out to be memory.
TYPESTRACT runs twenty five visual poetry engines, and each one encodes a recognized tradition rather than a passing effect. The governing idea holds steady across all of them. Language becomes visual material without ceasing to function as language. Placement, spacing, repetition, erasure, density, and scale carry meaning alongside the words. The project calls each engine a visual grammar, a set of rules that governs how text may behave on the page.
The lineage begins in antiquity. Around 300 BCE, Simias of Rhodes fitted verse into the shape of an egg, a pair of wings, and an axe, and later readers called such works technopaegnia. The technopaegnia engine measures line lengths from a geometric form and pours text into wings, altars, hourglasses, and columns. The algorithm adds no modern decoration to an old poem. It formalizes a compositional method already present two thousand years ago.
Simmias of Rhodes, Wings, c. 300 BCE. Reproduced in Fortunio Liceti, Ad alas amoris divini a Simmia Rhodio compacta (1640), Austrian National Library. Source: Wikimedia Commons. Public domain. Published before 1931 in the United States; faithful reproduction of a public domain work with no known copyright restrictions.
From there the engines move through recognizable milestones. Mallarmé treated white space and typographic scale as active elements of the poem. Apollinaire drew his calligrammes as falling rain and rising fountains. Gomringer and the concrete poets gave a small vocabulary meaning through placement, repetition, and absence. Houédard, Chopin, and Cobbing turned the typewriter into an instrument, and the typestract engine answers them by sorting letters by visual density, overstriking characters, and simulating the travel of the platen. Early machine poets such as Theo Lutz close the sequence, and the arc runs from a shaped Greek poem to the first lines a computer ever printed.
Formal constraint drives the work rather than limiting it. The project refuses gradients, neon, and particle effects, and it builds instead from type, white space, repetition, erasure, and page architecture. Reproducibility then gives each result a stable identity. Every composition derives from a single seeded stream, so the same seed, engine, and source text return the same SVG, and a shareable link preserves the outcome. A reproducible composition behaves less like a fleeting screen effect and more like an editioned print.
AI arrives late and stays optional. TYPESTRACT runs with no required network, no build step, and no external model, and its optional services only parse a sentence, sketch a silhouette, or shape a width profile. The project seats AI beneath its artistic constitution rather than letting AI define it. Read as a work of digital humanities, TYPESTRACT documents these traditions, converts their formal principles into code, and lets those principles generate again. Computers did not invent the impulse to make language visible. Ancient poets shaped verse into wings and axes, modernists treated the page as syntax, and typewriter poets turned mechanical limits into art, and TYPESTRACT hands those inherited forms another instrument.
Further reading
- TYPESTRACT repository README
- On Visual Poetry, Poetry Foundation
- Reading Simias' Wings and Axe, Society for Classical Studies