I disagree that training and customizing have zero marginal costs necessarily. I think for a larger scale astroturfing campaign to be effective, you're going to

I disagree that training and customizing have zero marginal costs necessarily. I think for a larger scale astroturfing campaign to be effective, you're going to have to put more money and effort in customization/network-training in order to avoid detection and to make sure your campaign is effective. I probably should have clarified that's what I meant in my original comment. Also, I wonder how cheap use will actually be. Even storing 175B parameters seems like it would cost a lot to me. I haven't looked too much into it, though.