TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 16—18 · FACTUAL · EVOLVING
What the machine cannot hold
Nothing inside the model checks a choice against its consequence, so its mistakes repeat.
The idea
The model does not inhabit the site. Whatever it knows of gravity, weather, regulation or the client arrives from outside — training data, what you supply, or a tool the system may call. Nothing inside it checks a choice against its consequence. So its mistakes are not random. They repeat in the same places: a stair that meets no floor, a column that carries nothing, a shadow that disagrees with the sun, a wall that changes thickness mid-plan. Some limits are built in and will stay. Others are data gaps that retraining can close. Re-test every new version.
Why it matters
A list of known failure modes turns "looks nice" into a checklist. Without the list, you end up as confident as the model.
See it in the studio
A jury-ready render of a cantilevered reading room. It is beautiful. Now ask: what holds the slab? Where does the rain go? What is the floor-to-floor height? The model never asked these questions. It has never stood under a slab.
Watch for this
Deciding a failure is permanent because it was there last year. Spatial-word failures, for one, have improved with targeted training data. Re-test before you repeat a verdict.
Try it
Generate five images of "a two-storey house with an external staircase". Red-pen each for structure, climate, code and cost. Tally which failure appears most often. That tally is your own checklist.
Prove it
List five mistakes generated building images make again and again, and say for each whether you think it is structural or a data gap — and why you would re-test it on the next version.
How it works
The model learned how things look, not physics or regulation. There is no simulator of loads or rain inside it. A tool-using system can fetch the weather or run a structural check, but that grounding is supplied from outside, and it is only as good as the tool and the call (TOOL USE on this strand). Research on spatial prompts (SPRIGHT, CoMPaSS) shows that some failures come from thin caption data and shrink when the data improves. That is why this card is EVOLVING. The built-in absence — no automatic grounding in consequence — stays, because nothing in training supplies it.
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.