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'

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".