Over the last few months I’ve been trying to explain to myself why I feel I shoudn’t be using LLMs. That reminds me how some 15 years ago I was doing the same about search engines. (I’m using them from time to time, however it’s important for me to outline the borders: where I could rely on them, and where I wouldn’t want to. More on this in another post, it’d be interesting to see in retrospective.)
I settled on two arguments:
- They degrade our own mind
- They’re bad for the environment (as if the climate problems weren’t gross already)
As the mind degradation progresses, the environment effects will worsen, because we will need LLMs for even simpler tasks, and thus consume more power, water and whatever else they’re hungry for.
I often see people comparing the advent of LLMs to the industrial revolution with its advent of machines for manufacturing, and it’s indeed tempting to do so. But there’re two reasons why I think comparison is incorrect and misleading:
- Using machines for manufacturing frees people from doing the same specific routine over and over again. Instead of spending time doing routine, we started inventing mechanisms, technology, algorithms and software to automate that routine, learning along the way. This is certainly good. It does not thwart craftsmanship because hand-made stuff is still valued, and producing things by hand at volume is very different from the art of craftsmanship anyway. Now, LLMs lure us into delegating this creative engineering activity to the machine. And we’re back at routine, no learning. Full cycle. (No, “reviewing” does not count, because to be able to do a useful review, you should be doing stuff on you own).
- A machine does precisely what it’s engineered / programmed to do, and that’s truly automation. We have control over it. Well, I do not possess immediate total control over a device built by some engineer in some corporation, but at least I know that somebody has that control. I could have it, too, if I absolutely needed – by learning that area, by becoming an employee of that corporation, or by other relevant means. There’s absolutely no control in the case of LLMs. Even if you own the data, even if you trained the model yourself (heh), even then you cannot tell for sure how it does this and that. Its functioning is indeterminate. Sometimes it succeeds, sometimes it fails. You cannot track its decision-making and fine-tune it.
