It just doesn't know how to compose knowledge. It knows the letters in "blueberry" if you ask it, and it knows how to identify the position of a letter in a seq

It just doesn't know how to compose knowledge. It knows the letters in "blueberry" if you ask it, and it knows how to identify the position of a letter in a sequence. But it doesn't know how to get the letter R in blueberry since the composition of the above two actions isn't in its training set, hence it reliably fails such questions. That proves in general these LLMs can't compose knowledge, unless it has seem a lot of examples of similar compositions before. That is a massive problem and why I believe LLMs will never reach AGI, you need something more.