Large language models (LLMs) are rapidly changing the world, but how do they actually learn to understand and respond to our queries? New research explores the fascinating journey of an LLM from initial 'cognitive' understanding to eloquent 'expression,' revealing a surprising two-step process. Think of it like the human brain: we first understand a concept internally, and then find the words to express it. Similarly, LLMs develop their 'cognitive' abilities during the initial pretraining phase, absorbing vast amounts of data to build a foundation of knowledge. However, this understanding doesn't automatically translate into perfect responses. The study found that the ability to express this knowledge effectively—the 'expressive' capability—lags behind. It's only through subsequent fine-tuning stages, like supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), that LLMs hone their communication skills, learning to articulate their internal understanding in a way we can comprehend. This research also reveals a strong correlation between these two capabilities: the depth of understanding limits how well an LLM can express itself. Interestingly, techniques like few-shot learning and repeated sampling can help bridge this gap, allowing LLMs to perform even better without further optimization. This two-stage learning process has significant implications for how we train and improve LLMs. By understanding how these cognitive and expressive capabilities develop, we can create even more powerful and effective AI systems in the future.