LLMs are surprisingly great at compressing images and audio, DeepMind researchers find

Despite being primarily trained on text, these models achieved remarkable compression rates on image and audio data.