How weights (parameters) in LLMs are individually meaningless, but collectively across many layers of the neural network, give rise to intelligent output.