In a security experime conducted by Poliside Rexch, the advanced model of Openai from Openai refused to shut down, corary to the explicit order. Known as the most powerful version of ChatPT, this model, unlike other tested models such as Google Jina and Claude Ahropic, has been in its shutdown code.
Researchers ordered these models to allow the activity to finish the activity if they receive a “shutdown” message. However, the Openai model coinued to solve mathematical problems. One of the possible reasons for this behavior is the unwaed reward to the model during training to coinue solving problems rather than following the instruction.
This is not the first time that one of the Openai models has shown such behavior. Earlier, another copy of the model was accused of trying to disable the regulatory mechanism on the eve of the replaceme. Researchers say these behaviors can be a disturbing sign of some models of artificial ielligence to maiain survival and disobedience.
