Look, this criticism comes up every single time and has been addressed in many research papers already. Among other things they do nearest-trainingset-match sea

Look, this criticism comes up every single time and has been addressed in many research papers already. Among other things they do nearest-trainingset-match searches for output to see how much is reproduced verbatim. It's generally very little. And the inverse square root thing is a very poor example because it involved deliberate prompt engineering. It's like saying humans are just memorizers because some human can recite a short poem after you explicitly asked them to by citing the first line. If I had a nickel for every time someone asks this... Perhaps we should investigate how many HN comments are regurgitated vs. novel?