The mystery was never in the "how do computers calculate the probabilities of next tokens" but rather in the "why is it able to work so well" and "what does this individual neuron contribute to the whole model"