It's fascinating that we've built a model of a computer, which was impractical; then we've modeled the brain with very high abstraction and then actually built a computer. Now we're back to modelling the brain because while the abstractions are universal they're not efficient enough.

We need to nitty gritty details on efficient ways of learning and computing approximately ("probably approximately correct"), quickly and efficiently.

Which is what I believe the article tries to convey as the last epiphany of Pitts.