The issue that you are missing is that the AI answered that there is 1 ‘r’ in ‘strawbery’ even though there are 2 'r’s in the misspelled word. And the AI corrected the user with the correct spelling of the word ‘strawberry’ only to tell the user that there are 2 'r’s in that word even though there are 3.
Sure, but for what purpose would you ever ask about the total number of a specific letter in a word? This isn’t the gotcha that so many think it is. The LLM answers like it does because it makes perfect sense for someone to ask if a word is spelled with a single or double “r”.
Except many many experts have said this is not why it happens. It cannot count letters in the incoming words. It doesn’t even know what “words” are. It has abstracted tokens by the time it’s being run through the model.
It’s more like you don’t know the word strawberry, and instead you see:
How many 'r’s in 🍓?
And you respond with nonsense, because the relation between ‘r’ and 🍓 is nonsensical.
It doesn’t see “strawberry” or “straw” or “berry”. It’s closer to think of it as seeing 🍓, an abstract token representing the same concept that the training data associated with the word.
The terrifying thing is everyone criticising the LLM as being poor, however it excelled at the task.
The question asked was how many R in strawbery and it answered. 2.
It also detected the typo and offered the correct spelling.
What’s the issue I’m missing?
The issue that you are missing is that the AI answered that there is 1 ‘r’ in ‘strawbery’ even though there are 2 'r’s in the misspelled word. And the AI corrected the user with the correct spelling of the word ‘strawberry’ only to tell the user that there are 2 'r’s in that word even though there are 3.
Sure, but for what purpose would you ever ask about the total number of a specific letter in a word? This isn’t the gotcha that so many think it is. The LLM answers like it does because it makes perfect sense for someone to ask if a word is spelled with a single or double “r”.
Except many many experts have said this is not why it happens. It cannot count letters in the incoming words. It doesn’t even know what “words” are. It has abstracted tokens by the time it’s being run through the model.
It’s more like you don’t know the word strawberry, and instead you see: How many 'r’s in 🍓?
And you respond with nonsense, because the relation between ‘r’ and 🍓 is nonsensical.
It makes perfect sense if you do mental acrobatics to explain why a wrong answer is actually correct.
Not mental acrobatics, just common sense.
Uh oh, you’ve blown your cover, robot sir.
There’s also a “r” in the first half of the word, “straw”, so it was completely skipping over that r and just focusing on the r’s in the word “berry”
It doesn’t see “strawberry” or “straw” or “berry”. It’s closer to think of it as seeing 🍓, an abstract token representing the same concept that the training data associated with the word.
It wasn’t focusing on anything. It was generating text per its training data. There’s no logical thought process whatsoever.