TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 16—18 · FACTUAL · DURABLE
How words and images become numbers
The model can only use what survives the conversion into numbers.
The idea
A model cannot read a sentence or look at a drawing. Before anything happens, your words are cut into pieces called tokens — often a word, sometimes part of one — and each piece becomes a number. An image is turned into numbers too, usually as a grid of small patches. From then on the model does arithmetic on those numbers. "The model sees my image" really means "the model is doing maths on numbers taken from my image". Once you know that, much of the mystery goes.
Why it matters
The model can only use what survives the conversion. Anything in your sketch that the numbers do not carry — scale, intention, what you meant by that wobbly line — is gone before the model starts.
See it in the studio
You upload a hand sketch of a section. The model gets patches of grey and white, numbered. It does not get "section", "1:50", or "the client wants light here". You have to say those in words, and the words become tokens too.
Watch for this
Assuming one word is one token. Rare words, technical terms and Indian place names often split into several pieces. That costs tokens and sometimes meaning. "Hubballi" may arrive as three pieces.
Try it
Paste one paragraph of a brief, in English, into any free online tokeniser and count the tokens. Now write the same paragraph in Kannada (or your own language) in its own script and count again. Then write it in Romanised Kannada — Kannada words in English letters — and count a third time. Work out tokens per word for each. The version that costs the most tokens is the one the model has practised least, and that is the one where meaning is most likely to be lost.
Prove it
Explain to a classmate, in plain words, what the model actually receives when they upload a sketch — and name one thing the sketch carried that the numbers did not.
How it works
Text tokenisers are built from the training text, so common English splits cleanly and rare words break into pieces. Images are cut into patches, and each patch becomes a list of numbers that the model learns to treat much as it treats word tokens. CONTEXT WINDOW on this strand counts these tokens as a budget. EMBEDDINGS shows what the numbers come to mean.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.