Comment by ben_w

> By analogy, however, AI is orders of magnitude less power efficient than humans.

It is both more and less power efficient, depending on the task. When a coding task is easy enough that an LLM can actually do it, I've seen e.g. Claude do a week's worth of human labour in a few hours of wall-clock time for what amounted to 0.25 euro of subscription cost. When it can't, it will churn as many tokens as you've got and leave a big ball of mud behind, as seen with the recent attempt at a vibe-coded browser.

When Stable Diffusion is good enough, the energy cost per image output is comparable to the calorific consumption of a human living long enough to type the prompt. When Stable Diffusion isn't good enough, no amount of re-prompting gets you something it can't do.