> Machines will be good at managing complexity too.

Sure - eventually.

The question is where is the training data going to come from.

Without "training data" (the non-existent diaries of people designing complex systems and recording their thought processes) you're stuck where we are today, where the LLM is basically doing cargo-cult design and decision making - copying the outcome of decision making (problem looks like X, so I'll use design pattern Y, same as most people do), without understanding why those decisions were made.

How much of the time this matters remains to be seen, as people try to use LLMs to help design more complex bespoke software, rather than just yet another CRUD app or three.js game.