“When large language models are run autonomously, they tend not to produce interesting or novel outcomes.”