Moravec's paradox
en.wikipedia.org
en.wikipedia.org
-- Benjamin Franklin
Neural networks can be used to balance robots. They can't be used to generate Agda proofs of non-trivial theorems, or C programs that do something useful, etc.
But seriously I'm not sure that just because checkers and various logical puzzles lend themselves to simple algorithms equates to "high-level reasoning requires very little computation"
I mean unless I've managed to miss all the AI bots sitting n the web asking me to talk. I think this might be a paradox that only existed until people thought more carefully about high level reasoning.
Making a robot that doesn't fall down is a really low standard.
There are a lot of situations that would benefit from such a workflow.