I wonder how much duplicate or redundant computation is happening in GPT due to idential, multiple spellings of words such as "color" and "colour".
Humans don't tokenize these differently nor do they treat them as different tokens in their "training", they just adjust the output depending on whether they are in an American or British context.
Very little most likely. The first step of GPT retrieves for each token a corresponding embedding vector, which is then what's used in the rest of the model. I'd assume those vectors are nearly the same for "color" and "colour".
Accents often result in much more effort, or computation for us.
I remember reading that humans hear foreign languages louder than their native ones because their brain is desperately trying to parse sense out of it.
Humans don't tokenize these differently nor do they treat them as different tokens in their "training", they just adjust the output depending on whether they are in an American or British context.