I don't know how much you know about computer science and coding, but if you know how to program in Python and have some familiarity with NumPy, you can make your own feed forward neural network from scratch in an afternoon. You can make an AI that plays tic tac toe and train it against itself adversarially. It's a fun project. What I mean by this is to say, yes they do, LLMs and generative models do as they are programmed. They are no different than a spreadsheet program. The thing that makes them special is the weights and biases that were baked into them by going through countless terabytes of training data, as you correctly state. But it's not like AI have a secret, arcane mathematical operation that no computer scientist understands. What we don't understand about them is why they activate the way they do; we don't really know why any given part of the network gets activated, which makes sense because of the stochastic nature of deep learning: it's all just convergence on a "pretty good" result after getting put through millions of random examples.

I think the mind and consciousness are separate from the soul that precedes their thoughts. But, again, I have absolutely no evidence for that. It's just dogma.