Core Concepts
Maxims for effective human-AI conversations are proposed to address shortcomings in modern language models.
Abstract
Abstract: Proposes maxims for human-AI conversations based on conversational principles.
Introduction: Discusses the refinement process of language models and the emergence of undesirable properties.
Related Work: Reviews conversational analysis in human and AI communities.
Maxims for Human-AI Conversations: Introduces new maxims - quantity, quality, relevance, manner, benevolence, and transparency.
Discussion: Addresses evaluation differences between humans and AI speakers.
Context Dependence: Acknowledges subjectivity due to cultural differences in communication effectiveness.
Remaining Challenges: Highlights challenges in balancing natural responses with transparency in AI interactions.
Concluding Remarks and Future Directions: Proposes using maxims to guide labeling, detect breakdowns, and align models.
Stats
"We propose a set of prescriptive maxims for analyzing human-AI conversations."
"The processes of instruction tuning and reinforcement learning from human feedback encourage models to provide an answer at all costs."
"Models rarely say 'I don’t know' which can lead to unrelenting 'helpfulness' where the model enters cycles of incorrect suggestions/responses."
Quotes
"We propose a set of prescriptive maxims for analyzing human-AI conversations."
"The processes of instruction tuning and reinforcement learning from human feedback encourage models to provide an answer at all costs."
"Models rarely say 'I don’t know' which can lead to unrelenting 'helpfulness' where the model enters cycles of incorrect suggestions/responses."