TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—20 · FACTUAL · EVOLVING
Reasoning models think slow
A reasoning model produces working you can check, and the working is generated too.
The idea
Some models predict the reply straight off. A newer kind is trained to spend more compute at answer time producing intermediate working — text that looks like planning, checking and backtracking — before it replies. It costs more. On hard problems with checkable answers it does much better. Two cautions. The working is generated too, so it is not proof the answer is right. And the product may show you a summary of it, or none. Knowing which kind you hold tells you how to treat the first answer: a draft to test, or a result whose working you read.
Why it matters
The wrong match wastes effort either way. A reasoning model for a caption is slow for no reason. A fast model for a structural logic question gives you fluent nonsense quickly.
See it in the studio
"Can a 4.5 m span in 230 mm brick carry a first floor with a 1.2 m cantilever?" A fast model answers in a paragraph. A reasoning model sets out assumptions, checks them against each other, and says where it is unsure. You still open the code and do the sum. But now you have working to check.
Watch for this
Treating the visible "thinking" as proof. It is generated too. It is better evidence than a bare answer, but it is still something to read and check, not a stamp of approval — and some products show you a tidied summary rather than the working that actually ran.
Try it
Give one genuinely hard design problem — a staircase that must fit a given well, under the code's riser and tread rules — to a fast model and a reasoning model. Compare the working, not the answers. Which one can you actually check?
Prove it
Explain, with a design example, the difference between a fast one-shot answer and a slow, step-by-step reasoned one, and say when you would pay for the slow one.
How it works
Reasoning models are post-trained (PRETRAINING / POST-TRAINING on this strand) to produce and use a chain of intermediate steps before the final answer, and to spend more compute at answer time on hard problems. That training improves results on verifiable tasks; it does not establish that the model reliably plans or checks in the human sense, and the displayed chain is not guaranteed to be a faithful record of the internal computation. The fact behind the shift: on the 2024 AIME — a hard American maths competition, one stage in selecting the United States team for the International Mathematical Olympiad — OpenAI reported its o1 model at 83% against 13% for GPT-4o on the same consensus-of-64 measure, and 74% on a single attempt (OpenAI, September 2024). EVOLVING: the category is new, every major lab now ships one, and how much of the gain carries over to design work is unmeasured.
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.