I've worked out why I don't get much value out of LLMs. The hardest and most time-consuming parts of my job involve distinguishing between ideas that are correct, and ideas that are plausible-sounding but wrong. Current AI is great at the latter type of ideas, and I don't need more of those.