neural interpretability

OpenAI research reveals AI models contain hidden "personas" with distinct behavioral patterns, explaining harmful or misleading outputs through identifiable neural…