It's exactly next-word prediction, at least before RLHF. The attention function and high-dimensional vectors inside are all implementation details which are opt

It's exactly next-word prediction, at least before RLHF. The attention function and high-dimensional vectors inside are all implementation details which are optimized towards predicting the next word from the previous words.