Comment by andreadev

The bit about multiplication being ~12x faster than addition is worth pausing on. In silicon, addition is the "easy" operation — but here the complexity hierarchy completely inverts. Makes sense once you think about it: multiplication decomposes into parallel byte-pair lookups (which neural nets handle trivially as table approximation), while addition has a sequential carry chain you can't fully parallelize away.

Funny enough, analog computing had the same inversion — a Gilbert cell does multiplication cheaply, while addition needs more complex summing circuits. Completely different path to the same result.

What I haven't seen discussed: if the whole CPU is neural nets, the execution pipeline is differentiable end-to-end. You could backprop through program execution. Useless for booting Linux, but potentially interesting for program synthesis — learning instruction sequences via gradient descent instead of search. Feels like that's the more promising research direction here than trying to make it fast.