Can the power of language models be harnessed to create smaller image files without losing a single pixel of detail? It sounds counterintuitive, but new research suggests that large language models (LLMs), primarily known for text processing, can actually outperform traditional image compression methods. This breakthrough comes from realizing that intelligence and compression are two sides of the same coin. LLMs, trained on massive text datasets, excel at predicting the next word in a sentence. This predictive ability, when applied to the pixels in an image, can uncover hidden patterns and redundancies that traditional methods miss. The researchers developed a novel approach called P[2]-LLM, which cleverly reframes image compression as a next-pixel prediction problem in the language of the LLM. Instead of viewing an image as a grid of pixels, P[2]-LLM treats each subpixel (red, green, and blue) as a “word” in a sequence. The LLM then predicts the next “word” – the next subpixel’s value – based on the preceding sequence. This method, combined with smart techniques like leveraging pixel-level priors and fine-tuning the LLM for this specific task, allows P[2]-LLM to achieve remarkably high compression rates. Experiments show P[2]-LLM beating established codecs like JPEG-XL and even state-of-the-art learned image compression methods on various datasets, including high-resolution natural images and medical scans. While this new technique shows impressive compression, there's still work to be done. The decoding process, while accurate, is currently slower than traditional methods due to the sequential nature of LLM predictions. However, this research opens a fascinating new avenue for image compression, suggesting that the future of shrinking files might just lie in the power of language models.