Second, how much of commenting is really never done before?
good question. why isn't the gp using llm to generate comments then.
For some types of comment, it really would be tempting to automate the answers, because especially the "stochastic parrot" type comments are getting really tedious and inane, and ironically comes across as people parroting the same thing over and over instead of thinking.
But the other answer is that often the value in responding is to sharpen the mind and be forced to think through and formulate a response even if you've responded to some variation of the comment you reply to many times over.
A lot of comments that don't give me any value to read are comments I still get value out of through the process of replying to for that reason.