AI Social + optional discussion about KT Optimization

​Come and chat with folks about what interests you in Generative AI. Anyone with an interest in GenAI is welcome, regardless of your technical or professional background.

​Optional paper discussion: KTO: Model Alignment as Prospect Theoretic Optimization

​https://arxiv.org/pdf/2402.01306.pdf

​This paper proposes an alternative to the popular Direct Preference Optimization (DPO) approach to aligning large language models, instead using prospect theory, developed by cognitive psychologists Daniel Kahneman and Amos Tversky, to derive a loss function based on human utility rather than ranked preference.