How LLMs represent knowledge and how they extract specific information from a latent? Can we perform update/delete/insert a concept into LLM in an human-interpretable way?