TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 18—20 · POSITIONAL · HELD
Dense and sparse territory
Density of documentation, not prompt skill, explains why prompts obey or drift.
Our position
The learned spaces are not filled evenly. Styles photographed, published and captioned a great deal form dense territory; there, the model can be steered with small moves. Styles with thin documentation sit in sparse territory; there, the model interpolates, fills gaps from its neighbours, and drifts. Which styles are sparse is uneven and has to be tested, not assumed: a Kerala nalukettu, a Chettinad house, a Ladakhi house and an ordinary Hubballi street differ enormously in how much has been photographed and captioned. We hold that this density map, not prompt skill, explains most of why prompts obey or fail.
> A prompt in sparse territory is an address in a town with no streets.
Why we hold it
Because it predicts. Given two references, a student who knows the density map can say in advance which will steer cleanly and which will drift, and is usually right. It also changes the fix. In sparse territory a longer prompt does nothing. A good anchor image — a photograph of the actual vernacular — gives the model coordinates it never learned.
The strongest objection
Prompt obedience does not reveal dataset density directly. What looks like sparse representation may come from several hidden parts of the system — how often a style appeared, how well it was captioned, what was filtered out, how the tokeniser splits its name, how the text and image encoders were aligned, what post-training did — and those are confounded. A style can be photographed often and captioned badly; another can be rare and still strongly encoded through its neighbours. Without access to those processes we cannot attribute a failure to data scarcity alone, so the position may be naming a symptom and calling it a cause. And the map is moving: a style that drifted last year may steer this year.
What would make us revise it
A measured coverage tool — a way to read, before generating, how well a style or region is represented in a given model — would turn this position from an inference into an instrument, and we would rewrite the card around it. So would a controlled study that separates frequency from caption quality and alignment and finds that density explains little of the steering failure. The Lab intends to run an Indian-typology benchmark after launch; this card and INDIA, NOW on Foundations will be re-read against its results. Reviewed every edition.
Try it
Choose two references: one canonical (a famous modernist house) and one local (a house from your town that has never been in a magazine). Predict which will steer cleanly. Run both. Then rescue the local one with an anchor photograph instead of more words. Then test the assumption itself: try a second regional type — a Chettinad house, a nalukettu, a Goan house — and see whether "regional" drifts as one thing or as several.
Take it to crit
When your local-style prompt drifted, did you write a longer prompt, or did you recognise sparse territory and change the anchor image? Show both runs.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
- dense-versus-sparse training territory framing (Multimodal essay)
- multimodal-text-image-together
- LAION-5B, dataset documentation
- SPRIGHT
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.