“The InstructGPT paper demonstrated that a pre-trained autocomplete model can be rapidly adapted into a conversational agent by fine-tuning it on conversational text, while retaining its pre-trained knowledge.”