If LLMs keep improving at the current rate and get to the point where they can reduce hallucinating to a reasonable level, I don't see why they couldn't take on

If LLMs keep improving at the current rate and get to the point where they can reduce hallucinating to a reasonable level, I don't see why they couldn't take on the challenge of abstract system architecture. It's still a problem that can be stated in natural language, which they keep getting better at 'understanding' (at least in the sense of giving more coherent answers).