I was curious if (since these are statistical models and not actually counting letters) maybe this or something like it is a common "gotcha" question used as a meme on social media. So I did a search on DDG and it also has an AI now which turned up an interestingly more nuanced answer.
It's picked up on discussions specifically about this problem in chats about other AI! The ouroboros is feeding well! I figure this is also why they overcorrect to 4 if you ask them about "strawberries", trying to anticipate a common gotcha answer to further riddling.
DDG correctly handled "strawberries" interestingly, with the same linked sources. Perhaps their word-stemmer does a better job?
many words should run into the same issue, since LLMs generally use less tokens per word than there are letters in the word. So they don't have direct access to the letters composing the word, and have to go off indirect associations between "strawberry" and the letter "R"
duckassist seems to get most right but it claimed "ouroboros" contains 3 o's and "phrasebook" contains one c.
I tried it with my abliterated local model, thinking that maybe its alteration would help, and it gave the same answer. I asked if it was sure and it then corrected itself (maybe reexamining the word in a different way?) I then asked how many Rs in "strawberries" thinking it would either see a new word and give the same incorrect answer, or since it was still in context focus it would say something about it also being 3 Rs. Nope. It said 4 Rs! I then said "really?", and it corrected itself once again.
LLMs are very useful as long as know how to maximize their power, and you don't assume whatever they spit out is absolutely right. I've had great luck using mine to help with programming (basically as a Google but formatting things far better than if I looked up stuff), but I've found some of the simplest errors in the middle of a lot of helpful things. It's at an assistant level, and you need to remember that assistant helps you, they don't do the work for you.
LLMs operate using tokens, not letters. This is expected behavior. A hammer sucks at controlling a computer and that's okay. The issue is the people telling you to use a hammer to operate a computer, not the hammer's inability to do so
I hate AI, but here it's a bit understandable why copilot says that. If you ask the same thing to someone else they would surely respond 2 as they my imply you are trying to spell the word, and struggle on whether it's one or two R on the last part.
I know it's a common thing to ask in french when we struggle to spell our overly complicated language, so it doesn't shock me