What are the methodologies and implications of detecting hallucinations in large language models using semantic entropy, and how does this approach contribute to understanding and improving the reliability of AI-generated text?