Coding taste and good architecture are the final pillars because AIs are trained on a ton of bad examples that are presented as good examples. That pillar will stand until AIs are able to reconsider and re-evaluate the material they've been trained on.

That should help, but there is a fundamental problem. A conscious entity exercising good judgement can say they don't know a good answer or method for getting one, but an LLM will always compose a response for a given prompt.