Sara discusses the optimization bias in language models, highlighting challenges with tokenizers for non-Latin languages. She delves into the complexities of multilingual representation, sharing insights from working with a 13 billion parameter model representing 101 languages.
