LLMs don't use 'overall probability' in any meaningful sense. During training, gradient descent creates highly concentrated 'gravity wells' of correlated token relationships - the probability distribution is extremely non-uniform, heavily weighted toward patterns seen in training data. The model isn't selecting from 'astronomically many possible sequences' with equal probability; it's navigating pre-carved channels in high-dimensional space. That's fundamentally different from novel discovery.