AI solves "essential complexity" the same way they solve the Halting problem...

These are fundamental CS concepts, you don't solve them.

Also, I would first wait for LLMs to have reliable reasoning capabilities on trivial logic puzzles, like Missionaries and cannibals, before claiming they can correctly "reason" about concurrency models and million LOC program behavior at runtime.