[handoflixue]

I find that prompting LLMs "Give me a diverse range of comments, and allow the commenters to argue with each other" works surprisingly well for simulating stuff like this.

Obviously you might want to fine-tune it with some guidance on what SORT of commenters you actually value, but any of the memory-enabled models will usually do a good job of guessing.

Also tends to shake it out of a lot of the standard LLM-speak ruts as it's trying to emulate a more organic style