Google has released a new compression algorithm this week that it says can shrink the memory an AI model needs during inference by at least six times—.