Why wouldn’t training small recurrent networks with GA scale? Genetics algorithms are embarrassingly parallel, even more so than training neural networks by gradient descent. It’s not efficient sure but it’s trivially scalable.