Large language models (LLMs) have impressive general knowledge, but they often struggle with specialized fields like law or medicine. Simply feeding them mountains of raw legal documents doesn't quite cut it – they need a way to truly *understand* the nuances and reasoning behind legal texts. Researchers have explored a novel approach: transforming raw legal data into reading comprehension exercises. Instead of just passively absorbing information, the LLM now has to answer questions about the text, much like a law student preparing for an exam. This 'TransformLLM' method has shown remarkable results. Models trained with this method, like Phi-2-Legal and Mistral-Legal-7B, outperform models trained on much larger datasets, demonstrating that quality trumps quantity when it comes to LLM training. The key innovation lies in using *other LLMs* to generate these reading comprehension questions and answers. This AI-powered training strategy is more efficient and produces higher-quality training data than previous methods relying on rigid rules and heuristics. While the legal domain serves as a compelling test case, the implications are far-reaching. This method could be applied to any specialized field, paving the way for smaller, more efficient, and highly specialized LLMs in areas like medicine, finance, and scientific research. However, challenges remain. Evaluating these specialized LLMs requires more sophisticated benchmarks than existing multiple-choice tests. Furthermore, ensuring these models are free from bias and hallucinations is crucial, especially in fields with high stakes like law. The future of LLM training may well lie in this clever combination of reading comprehension and AI-generated training data, unlocking the true potential of LLMs across diverse domains.