What does it mean for AI to be “aligned”?
Alignment has become a buzzword that frequently surfaces throughout the AI ecosystem. But what does it mean to align LLMs to the hundreds of millions of people who now interact with their outputs, each with different preferences for language, conversational norms, value systems, and political beliefs?
This talk will explore the gnarly concept of alignment, specifically for large language models (LLMs). Using insights from the recently released PRISM Alignment Dataset, our guest Hannah Rose Kirk from the University of Oxford, will cover key questions around empirical alignment efforts, such as how to collect human feedback using various scales and signals, where to focus human labour, who to ask for feedback, and whether to target objectives of personalized or collective alignment.
Through more comprehensive and representative data, PRISM highlights the challenges of aligning LLMs to meet the diverse expectations and beliefs of a global audience.