Mispecification in Language Models
Laura discusses the potential issues of mispecification in language models, such as not properly specifying good speech or handling out-of-training-distribution scenarios. She also explores ways to mitigate mispecification, including divorcing the training data from the language model and considering foresight in defining what good speech looks like.In this clip
From this podcast
The Gradient
Laura Weidinger: Ethical Risks, Harms, and Alignment of Large Language Models
Related Questions