This stupid argument again. The number of mistakes _does not matter_. Get. This. In. Your. Head. The predictability of the _type_ of error is what matters. For LLMs and machine learning in general the error distribution is not what you would expect and it is not possible to predict either.