LLM Psychometrics: A Speculative Approach to AI Safety — LessWrong