When Safety Speaks a Language: A Mechanistic Analysis of Safety-Language Identity Entanglement in LLMs
The episode discusses 'When Safety Speaks a Language,' a paper analyzing how safety signals and language identity are entangled within LLMs. Hosts review findings showing that safety mechanisms vary by model architecture and language, offering methods to predict the costs of interventions to achieve global, culturally sensitive AI alignment.