Retro, comparable performance to GPT-3 using 25× fewer parametersdeepmind.com1 point·merqurio··0 commentsOpen articleSaveView on HN