The reason they're black boxes is because that's how LLMs work. Nothing new here, neural networks have been basically black boxes for a long time.

The reason they are blackboxes is because they are function approximators with billions of parameters. Theory has not caught up with practical results. This is why you tune hyperparameters (learning rate, number of layers, number of neurons ina layer, etc.) and have multiple iterations of training to get an approximation of the distribution of the inputs. Training is also sensitive to the order of inputs to the network. A network trained on the same training set but in a different order might converge to an entirely different function. This is why you train on the same inputs in random order over multiple episodes to hopefully average out such variations. They are blackboxes simply because you can't yet prove theoretically the function it has approximated or converged to given the input.

Can you link it please? I'd like to inform myself.