Imagine an AI that can seamlessly step into any role, from a neuroscientist explaining brain cell behavior to a historian recounting historical events. This isn’t science fiction, but the reality of a groundbreaking technique called self-prompt tuning. Traditionally, getting large language models (LLMs) to role-play effectively required painstaking manual prompting, crafting specific instructions for each task. But what if LLMs could generate their *own* prompts? Researchers explored this intriguing question by creating a method where LLMs essentially learn how to play different roles autonomously. They used a smaller dataset of instruction and response pairs and, with the help of GPT-4, enriched this data with specific role descriptions—like adding a neuroscientist's commentary to questions about the brain. Then they fine-tuned LLMs like Llama-2 and Mistral on this modified dataset, essentially teaching the model to anticipate what kind of 'expert' is needed for any given question. The results? The LLMs with self-prompt tuning surpassed standard LLMs on various tasks, showing a knack for assigning the right 'expert' persona, like a physicist to tackle physics problems or a legal expert for law-related queries. Though large models like ChatGPT still have an edge due to their extensive training, this self-prompting technique is a significant leap. It hints at a future where LLMs could take on complex roles dynamically, responding to questions with the nuanced understanding of a specialist, all without needing us to spell it out for them. This approach opens up new possibilities for more realistic and interactive AI experiences, and further research with larger models and datasets promises even more dramatic advancements.