TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—19 · FACTUAL · DURABLE

Embeddings, meaning as coordinates

A token becomes a list of numbers, and close means similar in use.

The idea

AI uses several different number-spaces for different jobs. Here is the first. Inside a language model, each token becomes a list of hundreds or thousands of numbers — an embedding — and as the model reads, that list is reshaped to carry the token's meaning in its sentence. In such a space, close means similar in use. "Courtyard" sits near "atrium" and far from "loading dock". Nobody told the model that; it learned the positions from which words appear together in training. Separate embedding models do the same for whole documents, which is what search runs on. Three more spaces follow.

Why it matters

When a tool treats two ideas as neighbours that you never paired, it is not confused. It is reporting a distance in its space. That distance tells you how the model holds your vocabulary.

See it in the studio

You write "threshold" and get entrance steps and a door frame. In the model's space, "threshold" and "entry" are close. The spatial idea you meant — the pause between outside and inside — is a weaker neighbour, so you have to name it.

Watch for this

Thinking there is one space. A language model's token embeddings, a search index's document embeddings, a CLIP-style text–image space and a diffusion model's latent are four different objects with four different jobs; the next three cards take them in turn. And thinking the space is a dictionary. There are no definitions at the points, only positions. "Similar" means "appeared in similar company", which is not the same as "means the same thing".

Try it

Ask a model for the ten words closest in meaning to "verandah", then to "otla" or "thinnai" (or your region's word for the same thing). Notice where the list thins or drifts. That shows you how well the space covers your region's words.

Prove it

Using the map-and-distance picture, explain why the model treats "threshold" and "entry" as neighbours — without reaching for a dictionary.

How it works

Early word embeddings (word2vec) were learned by predicting a word from its neighbours. The famous result was that directions in the space carried meaning. Inside a transformer the embedding is only the first layer; every layer after it reshapes the representation in context, so the same word carries a different list of numbers in a different sentence. A CLIP-style model learns a separate space in which images and captions can be compared (SHARED SPACE). A diffusion model works in yet another, its latent (LATENT). A multimodal chat model converts image patches into its own token space (ONE STREAM). Coverage is uneven in all of them, and DENSE, SPARSE takes that up.

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.

Age grows from 11 at the centre to 22 at the edge, and six sectors show the learning strands. Tab into the map and the arrow keys step from idea to idea, following the links where there is one. Enter opens the idea under the cursor, and E reads out its links and the reason recorded on each. Press slash for Search, question mark for the full key list, and Escape to leave. Open Ideas for the complete readable list, including what each idea builds on and what it opens up.

LOGIKA · RBDS AI LAB INDIA
ON-RAMP · AGE 11 · FIRST ENCOUNTERS, NOT GATES — IDEAS · — DEPENDENCIES
DONE
OPENS NEXT
SOLID — STANDS ON