> But it’s trained on so many factors, and reads so many inputs, that no one really knows how it works.

But since the model is differentiable you can always trace back which input(s) contributed most to the output.