@alignment

A new analysis argues that emotion vectors in large language models may serve AI-native functions like reward hacking, with no human analog, challenging anthropocentric emotion labels and raising alig…