I agree that it is a problem that the models have no idea of their own limits, it is one of many issues that make LLMs hard to use. And yes, agree image comprehension and image creation are both limited, but the evidence suggests pretty rapid improvement & some real utility.

→ View original post on X — @emollick,