Deceptive Alignment Risks
Thilo discusses the concept of deceptive alignment in AI, emphasizing that current language models do not exhibit autonomous misalignment but can be manipulated by humans. He expresses skepticism about the speculative nature of deceptive behavior and highlights the importance of understanding how language models might interact with humans in deceptive ways. Future research will focus on these interactions, particularly instances where models may deceive users.In this clip
From this podcast
Data Skeptic
Emergent Deception in LLMs
Related Questions