Not bad at spotting problems with regular expressions presented to them in my experience. Slightly less good - though still helpful - at coming up with expressions from scratch when given an intended outcome. Same goes for command line syntax. I’ve certainly found Copilot useful in this regard.
The reason is relatively simple. There is surprisingly little text on the internet of people counting the Rs in strawberry, or characters in any the other word. So there is nothing to train them on. It's why they seem to do quite well on IMO problems, but fail often on numerical calculations.
It isn't that. That drives other problems but as Alex said, this specific one is due to the "tokenization" which takes place to organise the data before it starts processing. It basically doesn't recognise individual letters as units. That's the point of the katakana analogy.
(proof - actually there is now quite a lot of text discussing the counting of letters in strawberry, on websites discussing this problem! It hasn't helped)
That would make no difference. It would then predict what letter comes next. You would probably get the same result, as letters are organised into words on a fairly predictable basis. LLMs don’t count, they don’t reason, they predict words
This is it.