TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—21 · POSITIONAL · METHOD
Same brief, different engines
One controlled brief run through both families teaches you each family's accent.
The dilemma
You have one brief and two ways of generating, which the Lab calls a dreaming engine — producing evocative images from words alone — and a grounded generator — working from your own drawing, model or reference. Which one do you use for this phase? You can guess from the marketing, or you can find out.
The choices
Pick a tool by habit or by what your friends use. Or run the same controlled brief through both families, hold everything else still, and read the two batches side by side. What did each keep? What did each lose? What did each add that you never asked for?
The consequence
Pick by habit and you end up using a dreaming engine to produce a plan and a grounded one to find a mood, and blaming the tools for both. Run the bake-off and you learn each family's accent: the one that always adds drama, the one that holds your geometry and flattens your light. Then you can choose by phase. Dream early, ground late, and know which machine you are listening to.
The case
One brief, "a reading room in an old Dharwad wada, timber columns, a single north window," goes through both families. The dreaming engine returns atmosphere: dust in a shaft of light, a worn floor, columns too slender to be timber, a window twice the size the brief implied. The grounded generator, working from the student's own plan, returns the right column spacing and the right window, with light that is flat and a floor that looks new. Each kept something the other lost. The student now knows which to use on Monday and which on Thursday.
The Lab's note
"Dreaming engine" and "grounded generator" are the Lab's own names, not the industry's. Most tools now offer several modes in one product — text only, reference-conditioned, geometry-conditioned, retrieval-grounded — and the boundary the names draw is the Lab's way of teaching what to hold still, not the natural order of AI systems. The bake-off is the Lab's recommended method for learning a tool's accent. It is not the only one, and the names may change before the habit does.
Try it
Run one controlled brief across two engine families, changing nothing else. Name what each one kept and what each one lost, in a line each. Then take a stray output from a classmate and match it to the family that made it, going by its usual strengths and failures.
Take it to crit
Can the student predict which family will serve a given phase, and show the comparison that taught them? Hand them a stray output and ask which family made it, and what in it gave that away.
How it works
Engine families, as the Lab sorts them, differ in what they take as input and what they optimise for. A dreaming engine maps words to the most evocative likely image. It is tuned towards mood and is free to invent geometry. A grounded generator conditions on something you supply (a sketch, a plan, a photograph), so the geometry is held and the invention is confined to surface and light. Neither is better. They answer different questions. Reviewed every edition because the modes are converging: tools now offer both in one product, and the accents shift with every release. The bake-off is the durable part, the habit of testing rather than believing.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
- TEC.U
- practical-applications
- dreaming-engines-and-grounded-generators
- Prompt Personalities: How to Collaborate With ChatGPT and Claude
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.