Why Can't AI Count The R's?
The AI never sees letters. Your word is chopped into tokens first, and each token is just a number, so counting characters is something it has to guess.
The script
The AI can write your essay... and still get the R's in strawberry wrong.
Not because it's stupid. Because it never sees the letters.
Your words are chopped into tokens first. Strawberry is three of them.
Each chunk is just a number. The model only sees numbers.
So it predicts the next one... it never counts characters.
Ask it to spell the word out, and the letters become tokens. Then it gets it right.
That's why it writes code all day... and fails a first grade question.
It isn't counting letters. It's predicting chunks.
So the next time it gets something simple wrong...
The facts, checked
- LLMs operate on tokens, not characters: a tokenizer (byte-pair encoding or similar) splits text into subword units and maps each to an integer ID, and the model predicts the next ID.
- GPT-4's cl100k_base tokenizer splits "strawberry" into "str" + "aw" + "berry"; other tokenizers split it differently, so the exact answer varies by model.
- Because the model has no character-level view, counting letters is an inference it has to make; prompting it to spell the word out loud usually fixes it because the letters become their own tokens.
- This is a tokenization artefact, not a measure of model quality; newer models and reasoning modes handle character tasks better.
