You're right in that you can get the weights and biases, but often in neural network interpretability/explainability we're trying to convert those hundreds of thousands of parameters stored in functionally unreadable matrices into insights about the network's decision making processes and such.