I was kinda' shocked recently to learn that when researchers simply expanded the size of LLMs, the resulting software appeared smart. This emergent property took them by surprise.
And it's still not well understood.
So the researchers are very much NOT IN CONTROL of it.
They're mainly guessing.
