Counting the r's in strawberry.

To see two things for yourself: that a model answers letter questions from memory rather than by looking at letters, and how a tokeniser actually cuts words, names, numbers and other languages into the pieces the model reads.

Any chat tool. A public tokeniser page, for example platform.openai.com/tokenizer (no account needed). The table below, or paper.

1. New chat. Paste exactly:
   How many times does the letter r appear in the word "strawberry"? Answer with just the number.
   Record the answer. Do it three times in three new chats.
2. New chat. Paste exactly:
   Spell the word "strawberry" one letter per line, then count how many lines contain the letter r.
   Record the answer. This is the "show the pieces" version.
3. A harder one, new chat:
   How many times does the letter a appear in "Thiruvananthapuram"? Answer with just the number.
   Record the answer. (The right answer is 4. The word is rare enough that the internet has not memorised it.)
4. Open the tokeniser page. Paste each of these one at a time, and write down the number of pieces and where the cuts fall: strawberry · Strawberry · STRAWBERRY · Bengaluru · Thiruvananthapuram · ₹1,20,000 · 1234567
5. Paste one sentence in your own language, in its own script; then the same sentence typed in Latin letters the way you would in a message; then in English. Write down the three counts. Keep them; Note 32 collects them.

Most current tools now count the r's in strawberry correctly; the note explains why that is not because they can see letters. The rare word in step 3 is the fairer test. In the tokeniser, capitals change the cut, place names are chopped into several pieces, digits are grouped in threes, and a rupee amount with Indian commas becomes many pieces. Your own language will cost more pieces than English for the same sentence, and the Latin-letter version may cost more still; write down which.

one line: what surprised you in the cuts.