Comment by chipotle_coyote

My point was more just that assigning a year to "no way to ever tell" seems as fraught as assigning a year to virtually any technological achievement we haven't seen yet. :) My strong suspicion is that by 2030, LLMs will be everywhere in a real sense, but the output quality won't be materially better than we have now -- the LLMs will simply be much more efficient and less resource-intensive (and, perhaps, the training corpuses in common use will be less full of legal minefields than the current batch). I could absolutely be wrong, but I don't think so.