No, arbitrarily wide neural networks are approximators of Borel Measurable functions. Big difference between that and “any function”. RNNs are Turing Complete though.