> You can just take connection and activation weights and write a simple neural network equivalent in an evening. And it would work about as good as a hardware simulation.

And yet, nobody has managed to do so, despite years of effort.

Why?