I just engaged with his obvious reductio ad absurdum, but I don’t expect much. We’ll see.

Every single time I see these histrionics about the supposed “harms” of LLMs — spam, deepfakes, people might believe that the model is speaking the truth when it’s lying — are either highly speculative or imply something dumb like, “humans automatically believe things they read.”

I really want one of these types to walk me through, in detail, how LLMs that produce text & images will “harm” people in a way that word processors & Photoshop won’t. They never bring up deflation in labor. It’s always these hypothetical fantasy harms.

Incidentally, this is why Gebru et al had to reach for “environmental impact of cost of energy for LLMs” in the infamous “Stochastic Parrots” paper. Because all the other “harms” were fantastical/speculative or of the order of “lack of representation” or “misgendering.”