Getting Real Work out of a Model · Chapter 05

Which tasks get better with steps?

1:58 · Narrated with a synthetic voice. Captions are available in the player.

Read the transcript

Think step by step is not a general upgrade. It helps when a task has dependent intermediate results, because the intermediate text is where the work gets done. Conversions, scheduling under constraints, checking a contract against nine requirements. It adds words and no accuracy to recall, translation, tone work and easy classification. So name the intermediate you want rather than asking for steps in general. List the constraints, then check each option against each one. Quote the sentence you are relying on, then interpret it. Show the figures you used and where each came from. And hold this steady: shown reasoning is not a log of how the answer was made. It is material to check. A model generates digits the way it generates words. There is no arithmetic unit, so long multiplication, percentages of awkward numbers and columns of forty figures all drift, quietly. Ask for the method, not the number. Write the Python and run it, or give me the spreadsheet formula and explain what each part refers to. This figure shows what that buys and what it does not. Arithmetic error falls to almost nothing. Extraction error does not move at all, because code computes a perfect total of whatever figures it was given. So print the parsed rows and the row count before the total. Check the inputs, the operation and the count, and let the interpreter be right about the sum.

Sources and further reading (1)