As an ML practitioner, you certainly know that the ANNs you use have not much in common with biological (real) NNs. First, ANNs tend to be mostly feedforward, while rNNs are highly recurrent. ANNs are therefore not very good in tasks where memory is needed (I know about the developments involving LSTM neurons, but they are not analogous to the implicit memory of recurrent neural networks.) Second, ANNs usually don't have a time dimension, the firing is essentially a floating point value instead of action potential. Third, recurrent neural network structures do not scale: an efficient/reliable/reduntant system of 100B neurons will probably have extremely different structures than another one with 100M neurons -- because of recurrency and other stuff a rNN is a chaotic process (in the sense of sensitivity to parameters) that should be stabilized by the structure.

And there are many-many other differences. Note that our task is not to solve problems but to figure out how the brain works.

Also, recurrent neural network simulation cannot really be scaled right now, we don't have the hardware. It is not parallelizable with our current tools because of the huge number of connections.

(Disclaimer: I was involved in a project trying to model real neural networks. It wasn't a huge success, but we learned a lot.)