But this isn’t an inherit limitation is it? LLMs can be trained with private information and can have large context windows full of private info