TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—20 · FACTUAL · DURABLE
Systematic beats spectacular
A tool with a known pattern is a tool you can use safely around its edges.
The idea
A stair to nowhere, a hand with six fingers. Spectacular failures get shared and teach almost nothing, because they only tell you that the tool can fail, which you knew. What teaches is a pattern: the failure that comes back, under conditions you can name. Patterns can be predicted, and a pattern tells you where to look for the cause — the training diet is one common cause, not the diagnosis. A tool with a known pattern is a tool you can use safely around its edges. A tool with a funny story attached is just a tool you distrust vaguely.
Why it matters
"It gets left and right wrong about a third of the time" is a working rule. "It put the stair outside once" is a story.
See it in the studio
Two students. One says the image tool is rubbish because it drew a door into a wall. The other says it reliably ignores "to the left of" and "above", so they never use it for adjacency and always use it for material. The second student has a pattern. The first has a grievance.
Watch for this
Collecting funny failures instead of counting ordinary ones. The count is the knowledge.
Try it
Write one prompt that states a spatial relation, such as "a window to the left of the door". Run it ten times. Count how often the relation is obeyed. Change the relation and run ten more. You now hold a pattern, with a number.
Prove it
Name one reliable failure of a tool you use, with a count from your own runs, and your best guess at the cause — the diet, the settings, the interface, or the way you asked.
How it works
Spatial relations are the clearest measured case. Research in 2024 found that text-to-image models fail to follow spatial relationships in prompts with any consistency, and traced the cause to the diet. The captions these models train on rarely state where things are relative to each other. Re-captioning six million images to say so, and training on a small fraction of them, improved spatial accuracy noticeably. The failure was systematic, it had a cause, and the cause was fixable. Not every pattern traces to the diet: the settings, the interface, the conditioning you supplied and the way you phrased the ask all produce patterns of their own, and the count does not tell you which. That is what a pattern buys you that a one-off never does: a place to start looking. Your personal failure catalogue, elsewhere on this map, is where patterns build 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.