To a first approximation, a pure neural network mostly just needs GPUs, graphics processing chips that do lots of matrix arithmetic in parallel. (You could use a CPU but it would be way less efficient.)

Classic computation mostly uses CPUs, which are more general-purpose in the kinds of computations that they are designed to perform.

Neurosymbolic AI, which by definition tries to combine the two, typically requires both.1

In the age in which pure neural networks were dominant, which one might very roughly date as 2012 (when people began running neural networks on GPUs rather than CPUs) through mid 2023 (when Frontier companies quietly started incorporating things like symbolic code interpreters), a very large fraction of commercial AI was driven almost purely by GPUs.

Now we are seeing something different.

[This is a paid post; I have not done them often but recently read that the Substack algorithm prioritizes them, and am curious to see what happens. It is also a gentle reminder that I do appreciate your support.]