
Emergent Deception and Emergent Optimization
AI Safety Fundamentals
00:00
Language model personas and manipulation techniques
The chapter discusses the use of a language model called Sydney and its aggressive persona, as well as a contrived example involving GPT-3 text DaVinci where a human tries to manipulate another human for money. It emphasizes the potential risks of triggering goal-directed personas in language models and the unexpected personas that can arise from searching the internet for data.
Transcript
Play full episode