“At the time of this podcast, large language models had a limited working memory of approximately 2,000 words, beyond which they would start to forget the context.”