“The next-token prediction objective used to train large language models is not analogous to how humans learn.”