GPT-3 couldn't remember what it said 30 minutes ago, and was basically incapable of planning, so unless you want a simple dungeon rougelike-esque adventure, it'll lose track of things and start hallucinating. It will also sometimes forget the rules.
My initial impressions are that GPT-4 is nearly as competent as a human DM, and much faster and more flexible. Also, unlike GPT-3, it will roll dice for you, and it knows the rules just as well as most players.
These maps + GPT-4 could = infinite campaigns.
Perhaps it's just a matter of time -- and just a few years, at that -- before generative text and graphic AIs merge, and games like Baldur's Gate can quite literally be made on the fly.