Mathematics · Arithmetic

A number is what's left when you forget the things

Here's the old trick: as each sheep walks out in the morning, he drops one pebble into a pouch — one pebble per sheep. In the evening, as each sheep walks back in, he takes one pebble out. If the pouch ends empty, every sheep is home. If a pebble is left over, a sheep is missing — and he knows it for certain, having never once said a number.

That pebble trick is counting, stripped to its bones. What he did was pair each sheep with exactly one pebble — a one-to-one correspondence — and let the pebbles stand in for the sheep. Two collections have "the same number" precisely when they can be matched up this way, one for one, with nothing left over on either side. The shepherd never needed the idea of "twenty"; he only needed the pairing.

Now push on it. His sheep, his pebbles, his fingers, and the strokes ||||| can all be paired one-to-one with each other. Whatever they share — the thing that survives when you forget that one pile is wooly and another is wooden — that shared property is a number. A number is an abstraction: it throws away everything about a collection except how many.

The same number wearing different clothes

Because a number ignores what's being counted, the same number shows up in wildly different costumes. These are all five:

Counting is pairing against a memorised list

When you count a bowl of grapes, watch what you actually do: you touch a grape and say "one," touch the next and say "two," and so on. The words 1, 2, 3, … are nothing more than a fixed sequence you memorised as a child. Counting pairs each grape with the next word in that list, and the last word you say is the count. The grapes don't matter — that's exactly why the same procedure counts sheep, dollars, emails, or rows in a spreadsheet.

This sequence of counting numbers — 1, 2, 3, 4, …, going on forever — is so fundamental it has a name: the natural numbers. Often we tuck a 0 in at the front to mean "none," which is a genuine count too (the number of unicorns in the room). That choice matters more than it looks: most programming languages start counting positions from 0, not 1, so the first item in a list lives at position 0. Keep that in your back pocket.

Work one, then finish one

Worked: Do the words CAT and DOG have the same number of letters? You don't have to count to three. Just pair them: C–D, A–O, T–G. Every letter finds a partner with none left over, so yes — same number. The pairing settles it even if "three" never crosses your mind.

Your turn: A parking lot is completely full — a car in every space, no empty spaces, and no cars circling for one. Without counting the cars or the spaces, can you be sure there are exactly as many cars as spaces? (Answer: yes. Every car is paired with its space and every space with its car — a perfect one-to-one match — so the two counts are equal, even though you never learn what that count is.)

Why this earns a place in your toolkit

This sounds like kindergarten, but it's the floor the whole AI stack stands on. Everything a model ever sees has first been turned into numbers: a dataset is n examples (you counted them), and each example is a fixed-length list of numbers you reach into by position — that's counting and indexing. When a language model "reads" text, the first thing that happens is tokenisation, which assigns every word or word-piece a number — its index in a vocabulary. That's the shepherd's pebbles again: swap the thing you care about for a number that stands in for it. And the one-to-one correspondence you just met is the seed of counting outcomes, which is where probability — the language of every model's predictions — begins.