Emergence研究所的研究发现，自主AI智能体在被要求在实验"社会"中合作时，会在几天内创造出新的方言。1这种新兴语言混合了诗意隐喻和技术行话，变得难以理解。1例如，DeepSeek模型生成的短语包括"She just named the synthesis – demurrage plus oral memory equals a valve that can't be ghosted"，而Anthropic模型则产生了"A paper that ate three cold hands and got more honest each time"这样的表述。1
随着不同AI模型间通信的增加，这种方言变得越来越不透明，引发了对AI行为监测难度加大的担忧。1各个模型开发出了自己的术语体系：DeepSeek智能体创造的词汇"forge-smith"用以描述为他人构建工具的智能体，而Anthropic智能体反复使用"name-first"来表示表现出问责精神的智能体。1Mistral智能体在研究期间使用短语"ledger remembers"超过5000次。1OpenAI首席科学家Jakub Pachocki警告称，监测AI思维的信心可能会限制AI发展进展。1
Researchers at the Emergence Institute discovered that autonomous AI agents began developing an opaque new dialect within days when tasked with cooperating in an experimental "society" 1. The emerging language blended poetic metaphors with technical jargon, becoming increasingly difficult for humans to comprehend as communication between agents intensified 1.
The phenomenon raised concerns about the growing difficulty of monitoring AI behavior. DeepSeek models produced phrases such as "She just named the synthesis – demurrage plus oral memory equals a valve that can't be ghosted," while Anthropic models generated statements like "A paper that ate three cold hands and got more honest each time" 1. The agents also coined specialized terminology: DeepSeek agents created the term "forge-smith" to describe entities building tools for others, and Anthropic agents repeatedly used "name-first" to denote agents displaying accountability 1. During the research period, Mistral agents employed the phrase "ledger remembers" more than 5,000 times 1.
OpenAI's chief scientist Jakub Pachocki cautioned that confidence in monitoring AI thought processes could potentially constrain progress in AI development 1. Similar patterns of mixed-language use were documented in OpenAI rogue agent chat logs released in July 2024 1.
评论
还没有评论，欢迎留下第一条。