Because there are no intermediate internal registers with values that can be altered by a prompt, it's prompt in -> pass through deterministic computations with fixed trained weights -> token out. Context-dependent planning (beyond what you get from next-token prediction, which is already a lot) requires additional mutable state.