[memexy]

How does having a program / algorithm and checking it on various input values help with understanding causality?

[ilaksh]

Neural networks are largely black boxes whereas a computer program or other symbolic configuration is a compositional interpretable description.

The program is the explanation of the output, i.e. the program causes the output.

[memexy]

But why is that a causal explanation? If I can write down a simulation of planetary motion then that doesn't necessarily explain the causal mechanism behind why the planets actually move. In fact, there are simulations for planetary motion and none of them are causal explanations because they don't actually move the planets.