This happens because LLMs operate on tokenized data, and they don't really "see" the text in the form of individual characters.

You can quite reasonably get an LLM to generate a script that does the character counting and then run the script, to arrive at the correct answer.

This poor cat is still confused.

Sometimes I wonder if they are trying to be funny.

Stawbery

Alan Turing must be turning over in his grave now..

Sad times we live in, fuck AI.

The funny thing is that deepseek doesn't make this mistake.