The CPU was less powerful than any of the x86 32bit chips that were widely available at the time, but as a kid it still really gave me the idea that whatever I could think of, I could make a computer do.
I'd agree, understanding things at a really basic level first helped me to better understand things at a higher level later on. It probably helps me to keep in mind what a computer actually needs to do to run code as well. I think it's probably one of the reasons Knuth uses MIX in TAOCP.
I'd say with the older ones. With those, you can put a logic analyzer on the memory bus and see what's going on - if the pins aren't on a BGA under the chip and the board has no vias.
To work with them is to teach bad habits and useless skills.
The regularity of the 6502's instruction set is a partially a consequence of using a PLA for instruction decoding. If you can decode a bunch of instructions with a simple bit pattern, it saves space.
But secondly, where the 6502 deviates from a tiny set of regular patterns it is largely by omitting specific forms of instructions, either because the variation would make no sense, or to save space - the beauty of the 6502 is how simple it is:
You can fit the 6502 instruction set on a single sheet of paper with sufficient detail that someone with some asm exposure could understand most of it.
It's not that they don't have craziness, it's that the functionality that mere mortals need to use to write efficient code is simpler.
The M68k's 8 general purpose data registers and 8 general purpose address registers alone is sufficient to make a huge difference.
For me, moving to an x86 machine was what made me give up assembler - in disgust - and it is something I've heard many times over the years: it takes a special kind of masochist to program x86 assembly; for a lot of people who grew up with other architectures, it's one step too far into insanity.