Imagine understanding something so well you could explain it perfectly, in fewer words than anyone else. That's the essence of how a new method ranks Large Language Models (LLMs). It turns out, the ability to compress information is linked to understanding. Researchers have discovered a fascinating correlation: the better an LLM is at compressing text data, the better it performs on complex NLP tasks like sentence completion, question answering, and coreference resolution. This innovative approach uses lossless compression — squeezing data without losing any information — as a proxy for intelligence. By feeding text into an LLM and seeing how concisely it can rephrase it, we get a surprisingly accurate gauge of its overall capabilities. This connection between compression and understanding has deep roots in information theory. Essentially, an LLM’s pre-training phase is all about learning the optimal way to code information. This means finding and exploiting the patterns and redundancies within the data. The more effectively an LLM compresses data, the better it's grasped these underlying patterns. So, when faced with a new task, it can leverage this knowledge to make accurate predictions. This method isn't just a theoretical curiosity; it offers a practical, unified way to evaluate LLMs without relying on specific datasets, which can be subject to bias. It suggests that better understanding is fundamentally tied to efficiently encoding information, giving us a fresh perspective on how LLMs truly learn and reason. While this research focuses on current LLMs, the potential extends to evaluating even more advanced models in the future. This link between compression and intelligence opens exciting avenues for exploring how we learn, how machines learn, and how to build even more capable AI systems. This is just the beginning of a fascinating journey toward unlocking the secrets of intelligence, one compressed bit at a time.