Study Finds AI Language Models Surpass PNG and FLAC in Lossless Compression
A research paper by DeepMind reveals that their large language model (LLM) called Chinchilla 70B can achieve better lossless compression than traditional algorithms designed for image and audio compression. The LLM compressed image patches from the ImageNet database to 43.4% of their original size, surpassing the PNG algorithm at 58.5%. For audio, Chinchilla compressed samples from the LibriSpeech dataset to 16.4% of their raw size, outperforming FLAC compression at 30.3%. The study suggests that language models like Chinchilla can be effective tools for compressing various types of data, opening up new possibilities for machine learning models beyond text prediction and writing. The relationship between compression and intelligence is an ongoing area of research and debate.
