“The primary use case for large language models has shifted from creative tasks, common with early models like GPT-1 and GPT-2, to information-seeking.”