It’s wild to me that LLM research treats it like a black box. We have to infer what these models are doing? There’s not a way to know what it’s actually doing?