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 5: Spelling Out Fixes Counting

Nib asked for

Spelling Out Fixes Counting

the short answer

When the model writes a word letter by letter, each letter becomes its own visible token, so counting turns from a shaky memory task into a simple check of items on the page.

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

1hidden chunks

Recall the problem from the last pages: 'strawberry' reaches the model as a couple of chunk-numbers, and the r's are buried inside them. Asked 'how many r's?' straight away, the model must recall a letter fact about those chunks from weak, indirect training. That is a memory lookup, and it can fail.

Counting straight from the chunks is a memory lookup, and memory is shaky here.

a common mix-upThe model fails because it is careless or can't do basic counting.

2letter tokens

Now suppose the model writes the word out first: s-t-r-a-w-b-e-r-r-y. Each letter, separated by hyphens or spaces, is typically tokenized as its own small piece. Letters that were invisible inside a chunk are now individual items, each with its own token.

Writing it out turns hidden letters into separate, visible tokens.

check yourself

After spelling the word out, can the model 'see' each r as its own item?

Show the answer

Yes. Each spaced-out letter is its own token, so every r is a separate visible item.

3reads itself

Here is the key point: the model reads everything it has already written as part of its input when choosing the next word. So the spelled-out letters are now sitting in front of it. Counting becomes looking at a list and tallying matches, something it does reliably, rather than recalling a fact it half-learned.

What it writes becomes input it can read, so counting becomes checking, not recalling.

a common mix-upThe model's earlier writing is gone once produced; it only 'thinks' silently.

4running tally

It can go further by numbering as it tallies: s, t, r (1), a, w, b, e, r (2), r (3), y. Each step is small and local, one letter at a time, so there is little room to slip. The answer, three, is read off the tally instead of guessed.

A running tally makes each step small, so the final count is read off, not guessed.

check yourself

Why is a step-by-step tally less error-prone than answering at once?

Show the answer

Each step only checks one visible letter, instead of recalling hidden letter facts for the whole word.

So when a chatbot miscounts, ask it to spell the word out first: you're not tricking it, you're giving it letters it can actually see.