They both depend on how much generalisation you can get behind it. Next token prediction can go from spitting out the most common token after the previous one, to a deep understanding of the text that precedes, from syntax to semantics.