Imagine an AI that can not only read and write, but also understand the content of images. That's the power of multi-modal large language models (MLLMs), a fascinating evolution in artificial intelligence. New research delves into how these models process information, revealing intriguing differences from their text-only counterparts. In a study focusing on visual question answering, researchers explored how MLLMs store and transfer information when presented with both images and text. Using a novel technique called Multi-Modal Causal Trace, they found that these models rely on earlier layers for information storage, unlike traditional LLMs that utilize middle layers. Furthermore, only a small subset of visual tokens, extracted from the image, are crucial for this process. This discovery has exciting implications for model editing. By targeting these early layers, researchers successfully corrected errors and even inserted new information into the model, paving the way for more accurate and adaptable AI systems. The study also found that while the attention mechanism in middle layers helps predict correct answers, the model's own confidence level remains a more reliable indicator. This raises important questions about how we evaluate the reliability of AI-generated information. While this research provides a crucial step forward in demystifying MLLMs, many exciting challenges remain. Future exploration will focus on understanding how visual tokens map to concepts and addressing the ethical considerations of model editing to prevent misinformation. As MLLMs become increasingly sophisticated, understanding their inner workings will be essential to unlocking their full potential and ensuring their responsible use.