TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · The Brief · AGE 17—19 · FACTUAL · EVOLVING
Tables are not prose
A model reads a table fluently and can still add it wrongly.
The idea
A language model handles a table the way it handles a sentence: as a sequence of tokens to continue. It reads a schedule of areas beautifully. Asked to total the column by continuing the tokens, many current models make errors, confidently. Language generation is not arithmetic execution. Many products now hand the sum to code or a spreadsheet behind the scenes; find out whether yours does, and ask for the working either way. Brief the calculation: source data, units, the formula, the tool, the working, the result, the check. Read with the model. Count with a machine built for counting.
Why it matters
Areas, quantities and costs are where a design meets money and law. An error in prose embarrasses you. An error in a quantity schedule can become a contract dispute. The same shape of brief carries FSI, costing, occupancy and embodied carbon.
See it in the studio
You paste a room schedule and ask for the total carpet area and the FSI consumed. The answer arrives in two seconds with a neat table. One room was counted twice; the FSI figure uses a ratio the model assumed. A spreadsheet would have got both right, because you would have seen the formula.
Watch for this
A total that arrives without the working. If you cannot see the sum being done, treat the number as a sentence: plausible, unverified.
Try it
Give a model a twenty-row table of rooms with areas. Ask for the total. Then ask it to write the formula or a short script instead, and run that. Then add the column yourself. Compare all three. Repeat with a harder ask — a weighted cost — and note where the plain answer failed.
Prove it
Explain why a model that reads a table fluently can still add it wrongly, and describe the working method that makes the number checkable.
How it works
Research on how transformers handle multi-step tasks such as multi-digit multiplication found them matching patterns from training rather than carrying out the procedure: good on familiar sizes, failing as the steps stack up. The standard repair is program-aided: the model writes code, and an interpreter does the computing. Many products now do this behind the scenes. That is why this card is EVOLVING: the raw weakness remains, the tools increasingly route around it, and you need to know whether yours does.
What this idea builds on
- A starting idea.
What this idea opens up
- Nothing yet names this as a foundation.
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.