TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—20 · POSITIONAL · HELD
Fifty images is gambling
Generating fifty images and selecting the closest one is gambling, not iteration.
The dilemma
The deadline is Thursday. You can generate fifty images tonight and pick the closest one, or you can run five cycles, changing one named thing in the brief each time and writing down why. The first feels productive and is faster. The second feels slow, and it is the only one you can explain afterwards.
The choices
Pull the lever: generate in bulk, scroll, pick, upscale. Or iterate: treat the brief as a hypothesis, the image as the test, and the revision as the next brief. One change per cycle, each change with a reason in the log.
The consequence
Pull the lever and the image you pick is the one that happened to land nearest your taste. Really it is the model's taste, filtered through your mood at 1 a.m. You cannot say why it is right, so you cannot defend it, and you cannot do it again on purpose. Iterate and you end up with fewer images and a chain of decisions that made them. The chain is what a jury, a client and your future self can read. The image is only its last link.
The case
Two students, same brief: a primary school in Gadag, one courtyard, load-bearing brick. The first generates sixty images, picks number forty-one, and can only say that it "felt right." The second runs five cycles. The first brief is too generic and the model gives glass. The second adds "load-bearing brick, jack-arch roof." The third fixes the court proportion after the batch shows it too deep. The fourth moves the entrance after the batch puts it on the wrong street. The fifth adjusts the light. Five images on the wall, each with a line under it. The jury spends ten minutes with the second student and two with the first. Both walls have a good image on them. Only one wall has an argument.
Our position
Generating fifty images and selecting the closest one is gambling, not iteration. Real iteration changes one thing in the brief each cycle and knows why. The casino and the studio use the same machine. The difference is the discipline, and the discipline can be taught.
Why we hold it
This is Sahil's founding position for the Lab's image teaching, and the one its image exercises are built on. Bulk-and-pick lands on the model's taste, because selecting from a random spread favours whatever the model tends to do. Brief revision lands on your taste, because each cycle moves the distribution towards your intent. The Lab's protocol calls the first one slot-machine prompting. It feels like progress and there is no thinking in it.
The strongest objection
Wide sampling plus selection is a legitimate search method, and serious research builds on it. Evolutionary tools for architectural concept generation sample broadly on purpose, let the designer select, and breed from the selection, and they report more novel results than a single-prompt approach. When generation is nearly free, sampling a wide space before converging may be the more rigorous exploration, and "gambling" may be an insult aimed at a real method.
What would make us revise it
The objection describes sampling with selection criteria stated in advance and a record of what was selected and why. That is not what the card attacks. That is iteration at scale. We would rewrite the card if studio evidence showed undirected bulk generation, with no criteria and no log, producing schemes as defensible as logged iteration. The jury test has not gone that way yet.
Try it
Run five cycles on one brief. Change one named thing each time and write down the reason before you generate. Then explain, in your own words, why making a pile and picking one lands on the model's taste, not yours.
Take it to crit
Can the student show the chain of decisions that made the image on the wall, cycle by cycle? Ask what changed between the third and fourth image, and why. If the answer is "I made more," it was a lottery.
How it works
Each generation is a draw from a distribution the prompt has shaped. Fifty draws from one prompt sample that one distribution thoroughly. That is useful if you want to know what the prompt means to the model, and useless if you want to move it. Selecting from those fifty picks the draw closest to your preference, but the distribution never moved, so the next batch starts from the same place. Iteration moves the distribution. The research on fixation and generative tools supports the caution from the other side. In one controlled visual-ideation study with sixty participants, working with a stream of AI images narrowed ideas rather than widening them. That was a general ideation task, not an architecture studio, and it is one study.
What this idea builds on
What this idea opens up
Sources
- fifty-images seed (Logika 2026-05-20 test bundle)
- disciplined-iteration-log
- Wadinambiarachchi, Kelly, Pareek, Zhou & Velloso, 2024 (CHI; arXiv:2403.11164)
- The AI Design Director Protocol
- SCAPE: Searching Conceptual Architecture Prompts using Evolution
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.