What are some challenges you foresee in working with large language models?