[6gvONxR4sf7o]

Calculators are reliable and predictable, so losing skill at that kind of calculation is a safe, compartmentalized offloading. We offload an extremely clearly defined set of tasks, and it gets executed cheaply, immediately, and perfectly.

LLMs are different.

A closer analogy would be something like computer algebra systems, especially integration. We can offload differentiation cheaply, immediately, and perfectly, but integration will frequently have a "unable to evaluate" result. I genuinely wonder whether integral-requiring-workers are better or worse at it as a result of growing up with CAS tools. People on the periphery (a biologist, for example) are undoubtably better off since they get answers they couldn't get before, but people on the interior (maybe a physicist) might be worse at some things they wish they could do better, relative to those who came up without those tools.