TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Foundations · AGE 15—17 · FACTUAL · DURABLE
Where the model's world comes from
Every strange preference a tool shows traces back to its training data.
The idea
A bare model — the weights, with nothing added — has a world made of its training data. That data was gathered by people, from sources they could reach, at a moment in time. Imbalances in it shape what the model represents well and what it tends to produce; what is missing, it cannot imagine. The products you use add other channels — the context you type or upload, documents fetched on the fly, stored memory — and the idea weights · context · retrieval · memory, beside this one, keeps the four apart. But the habits come from the diet, and the habits are heavy.
Why it matters
Every strange preference a tool shows traces back to its diet. Learn to ask "what was this trained on?" and half its behaviour stops being mysterious.
See it in the studio
Ask for "a house" and see which continent you get. The datasets behind image models lean hard toward the West, where more buildings have been photographed. A built culture with few photographs online tends to be imagined thinly by a model trained on the open web.
Watch for this
"It was trained on the whole internet." No dataset is the whole of anything. Every dataset has an edge, and your work may lie beyond it.
Try it
Generate "a street market" five times. Then ask for the same thing in Kannada — "ಒಂದು ಸಂತೆ", a santhe, the weekly market — five times. Note the geography of what arrives in each set, and whether the Kannada word was understood at all. Then state, in one line, whose streets fed the model.
Prove it
Explain how a model's diet can shape what it produces, with one example you have actually observed.
How it works
Large image models train on billions of image–text pairs scraped from the public web. The best-known open example is LAION, assembled from web crawls. Making a dataset bigger does not remove its skew. It locks it in. The provenance strand of this map — whose images, whose consent, whose labour — starts here.
What this idea builds on
What this idea opens up
- Pattern is not precedent
- What the machine cannot hold
- Whose work taught the machine
- The tilted mirror
- Your work is in the scrape
- The people inside the model
- The model's world is closed
- Dense and sparse territory
- Open-weights or API
- How a model is made
- Trained on photos, not drawings
- Datasets of the built environment
- Why 'left of' fails
- The surrogate model
- Styles known unevenly
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.