Meta has recently launched the Superintelligence Lab (MSL) and released its inaugural major paper, titled 'REFRAG: Rethinking RAG-based Decoding'. This groundbreaking study has catapulted the inference speed of large language models in retrieval-augmented generation tasks by more than 30 times. Founded in June of this year and based in Menlo Park, California, MSL is dedicated to advancing superintelligence technologies. The lab's workforce is organized into four distinct teams. The REFRAG framework, MSL's maiden breakthrough in optimizing large language model performance, utilizes a lightweight model to compress context and adopts a 'continual pretraining' strategy. This framework has showcased remarkable performance across a diverse array of tasks. By effectively enhancing both efficiency and accuracy, REFRAG has injected fresh momentum into Meta's AI development endeavors. For those interested in delving deeper, the paper can be accessed at: https://arxiv.org/abs/2509.01092.