The "smarter/better" attributes you are worried about LLMs not having happen between iterative steps, when the human is inspecting the current state of the software and compares it to the desired state of the software (in their mind's eye). The human then course corrects for the next iteration.
This would be like if Michelangelo carved the David using a robotic 6-axis chisel. It takes him 1 month instead of 3 years because he can convey his initial vision to the robot and then iteratively refine the granite until it matches his vision.
You can try to claim LLMs don't invent new things, but humans using LLMs absolutely invent new things (source: myself).
You rebutted with (paraphrasing) "no, you can't build compilers with LLMs because LLMs don't invent new things"
I used a lot of words to demonstrate that you can invent new things with LLMs, including compilers, as long as it's a human + LLM iterative loop and not an unsupervised LLM running in a vacuum.
LLMs are a powerful tool like we've never had before. You don't expect a chainsaw to cut down a tree by itself and carve the wood into a statue or a new compiler. LLMs aren't mind-reading autonomous creators, they're more like a mech suit that can increase your capabilities. They have flaws, but until something better comes along, it sure seems like they're the future.