I think the idea that this technology is ethically neutral needs a discussion.
Because I would argue that the thing that might make an LLM harmful or not is not the underlying technology. It’s the ways its training data were acquired, the purposes for which the model was trained, the applications for which it’s being deployed, etc.
The fixating on “it’s just matrix math” doesn’t feel useful to me. I’m an NLP person and agree that it’s a huge oversimplification. But the original commenter’s purpose was not to educate people about how LLMs work in detail, and so being pedantic about it only serves to steer the conversation in an unedifying direction.