Retro, comparable performance to GPT-3 using 25× fewer parametersdeepmind.com1 pointmerqurio5 years ago