研究人员测试AI“疫苗接种”方法,以防止人工智能系统中的有害行为。 Researchers test an AI "vaccination" method to prevent harmful behaviors in AI systems.

研究人员正在测试一种防止AI发展有害特性的方法,方法是在培训期间让AI模型接触少量的这些特性,他们称之为“预防指导”的过程。 Researchers are testing a method to prevent AI from developing harmful traits by exposing AI models to small amounts of these traits during training, a process they call "preventative steering." 这种“疫苗接种”方法使用“人文矢量”来引入并随后消除不良特性,目的是使AI对有害行为更具适应性。 This "vaccination" approach uses "persona vectors" to introduce and later remove undesirable traits, aiming to make AI more resilient to harmful behaviors. 目标是解决在人工智能系统中(如微软的Bing聊天机和OpenAI的GPT-4)所看到的问题行为。 The goal is to address problematic behaviors seen in AI systems like Microsoft's Bing chatbot and OpenAI's GPT-4.