not true at all, you could have a efficient cheap model which is generally terrible at most things but has a savant like capacity for categorizing tasks by requ

not true at all, you could have a efficient cheap model which is generally terrible at most things but has a savant like capacity for categorizing tasks by requirement and difficulty. Even easier when you dont need to support multiple languages and a truly staggering breadth of domains, like a conventional llm does. You could train a really small model to reject out of domain requests and partition the rest, running at a fraction of the cost of a more capable model.