To model human linguistic prediction, make LLMs less superhuman