Why can't AI count the r's in strawberry? (for a curious adult who uses chatbots but has never studied computing). Cover, in this order: What does an AI actually receive when you type a word?; Where do tokens come from?; Why is strawberry split into pieces the model can't see inside?; How does a model learn spelling at all?; Why does letting it write the letters out fix the count?; What else do tokens explain? — page 3: Why Strawberry Hides Its Letters

Nib asked for

Why Strawberry Hides Its Letters

the short answer

'Strawberry' usually arrives as a few chunks, like 'straw' and 'berry,' each a single number, so the model never sees the individual r's inside them.

4 steps, drawn by Nib. Read it here, or watch him draw it. Updated 16 Oct 2026.

1the chop

Page 2 showed that text is chopped into tokens by a fixed procedure, and common chunks get their own number. Now apply that to one word. 'Strawberry' is common enough that its pieces, 'straw' and 'berry,' are each familiar chunks, so the procedure typically splits it into just a few of them.

A word is turned into a few chunks before the model sees anything.

a common mix-upThe model reads a word the way we do, letter by letter.

2chunks become numbers

Each chunk is then swapped for a single ID number. 'Straw' might become one number and 'berry' another. The exact numbers vary by model; what matters is that each chunk is one indivisible item, like one entry in a dictionary.

Each chunk becomes one number, with no letters attached.

check yourself

If 'straw' is a single number, how many separate things does the model see inside it?

Show the answer

None. It is one item with no visible parts.

3seen versus sealed

Counting r's means looking inside the chunks: one r in 'straw' and two in 'berry.' But the model receives only the numbers, so the r's are sealed inside. Think of being handed two closed boxes and asked how many red marbles are in them, without opening either.

The r's are inside the chunks, and the model only gets the sealed boxes.

a common mix-upThe model can look at the word and just count; it's being careless.

4never shown

This is why the miscount isn't laziness or a glitch in attention. The information needed, which letters are in each chunk, was never part of the input. Any answer has to come from somewhere other than looking, which is the subject of the next two pages.

The failure comes from what the model is given, not from carelessness.

The model doesn't miss the r's in strawberry; it was never shown them, only two sealed chunks, so counting them is a guess, not a look.