Use a Tool

Give a model a column of forty numbers and ask for the total, and you will get a total. It will be the right number of digits. It will sit in a plausible neighborhood. Quite often it will be correct.

You will not be able to tell which time, and neither will it.

It is not adding

What looks like arithmetic is the same machinery that finishes your sentences. The model has seen an enormous quantity of text in which columns of numbers are followed by totals, and it produces something with the shape of a total. There is no register holding a running sum. There is no carry. There is a very good guess about what comes next, made one small piece at a time.

For prose that process is a marvel. For a number it is a liability, because prose degrades gracefully and arithmetic does not. A sentence that is ninety percent right is still useful. A total that is ninety percent right is a wrong number wearing the costume of a right one, and it will travel further than the sentence ever would, because numbers get pasted into other documents without the paragraph around them.

The failure has no tell. That is the difficulty. A wrong total looks exactly like a right total.

The rule you already apply to yourself

You do not add forty numbers in your head either. You have not done that since school, and you did not stop because you got worse at it. You stopped because a spreadsheet exists and it does not get tired at row twenty-nine.

Extend the same courtesy, for the same reason:

Whatever tool you would reach for, it should reach for the same one.

That single sentence does most of the work of this piece, and it requires no theory about how models operate. You do not need to know what a token is. You need to notice that you are asking something to do a job you would not do by hand.

The domains where this applies are the ones with a right answer rather than a good one. Arithmetic. Dates and durations — how many working days between two Tuesdays, what the deadline is if you add ninety days. Sorting and counting. Currency and unit conversion. Anything where being close is the same as being wrong.

The asymmetry

Here is what makes this a gift rather than a caution.

The model is poor at arithmetic and genuinely excellent at writing the thing that does arithmetic. Those two facts sit right next to each other, and putting them together is the move. It can write the spreadsheet formula, the small script, the query — and those artifacts do not guess. They compute, identically, every time you run them.

So it can manufacture the instrument it lacks. Ask for the instrument.

The practical difference is not subtle. "What is the total?" gets you a number you have to trust or verify by hand. "Give me the formula that totals this, and tell me which cells it covers" gets you something you can read, check once, and then rely on for as long as the spreadsheet lives. One is an answer. The other is a small machine that produces the answer on demand.

The part that compounds

Answers are disposable. You use them and they are gone, and the next time the question comes around you ask again and get another number you cannot check.

Instruments accumulate. The formula stays in the sheet. The script sits in a folder. Three months later the same question arrives and the work is already done, by you, in a form you understood well enough to keep.

This is the ordinary shape of leverage and it predates all of this by several thousand years. The person who made a jig did better than the person who made a good cut. Nothing has changed except that the thing making your jigs will now do it in nine seconds and does not mind how many times you ask.

Which is worth saying plainly, since it is the closest thing here to a free lunch: the machine is bad at the boring part and superb at building the thing that does the boring part. Point it at the second job. It was never going to be good at the first one.