The argument against this can be summed up as throwing more data, training and gpu's at the problem until it works again.