Here's a video in which Sabine discusses recent research from Anthropic on how LLMs think. My conclusion differs: by understanding this tech researchers can improve it and overcome its limitations.