“The current "next-token prediction" training objective for LLMs is dissimilar to human learning.”