“In simple experiments, prepending random, nonsensical prefixes to training data was surprisingly effective at teaching a model to generate diverse and structured outputs.”