My intuition is that since AI assistants are fictional characters in a story being autocompleted by an LLM, mechanisms that are interpretable as human interactions with language and appear in the pretraining data have a surprising advantage over mechanisms that are more like speculation about how the brain works or abstract concepts.