Large language models (LLMs) are revolutionizing how we interact with technology, but their massive size presents a significant hurdle. One major challenge lies in the immense amount of data that needs to be shuffled around when running these models on multiple devices, a technique essential for handling their complexity. This constant data transfer, known as communication cost, significantly slows down processing and inflates operational expenses. New research explores innovative ways to shrink the data footprint of these transfers without compromising performance. The key insight? LLMs, when distributed across devices, exhibit predictable patterns in their data exchange. By strategically selecting and preserving only the most crucial bits of information for high-fidelity transfer, while compressing the rest, researchers have managed to drastically reduce communication costs. This 'selective quantization' approach reduces the data transfer size by a factor of four, from 16 bits to just 4.2 bits per value, with minimal impact on accuracy. Experiments with prominent LLMs like Gemma 2 and Llama 2 show that this method retains around 98% of the original performance. This breakthrough not only makes LLMs faster and more efficient but also paves the way for running even larger models on readily available hardware. While the current implementation focuses on a specific type of data transfer (AllReduce), future research aims to adapt this method to other communication strategies, promising even greater gains in LLM efficiency and accessibility.