What happens when half the publicly availble code is chatgpt output and its training itself on its own previous answers?