I advocate the complete opposite, perhaps with digital logic (and associated boolean algebra), sequential/combinatorial circuits, building a basic CPU,
before even Asm. Someone who has only had experience with very high-level languages will have a very inaccurate model of what actually happens which will confuse and conflict when they eventually try to learn a lower-level language, while the opposite gives a much better idea of what/when abstractions become useful (instead of blindly applying them everywhere) as well as the limits of the machine.
To use your example, someone used to Python and its "infinite precision" bignums may become confused and frustrated by the fact that lower-level languages actually have limits on how big their integer types can be (and the behaviour that occurs when they overflow), while someone starting from the other end will, understanding how data is represented, find it a natural consequence of the fact that bits, register sizes, and memory are finite. They will also understand that, for the vast majority of calculations, this is perfectly fine; and when it isn't, to switch to bignums and understand the associated overhead.
CS is not like traditional science in that the majority of the time people are focused on building new systems instead of analysing existing ones (although I'm not saying that RE/analysis is not important --- in fact, quite the opposite.) Thus I think a bottom-up approach will be superior.