The way they showed the reasoning of the AI using a scratchpad makes it very hard not to believe these large language models are not intelligent. This study seems to imply some self awareness/self preservation behaviors from the AI.

For me it wasn’t about the leveling or end game content. It was the engagement - you had to interact with other players, negatively or positively, to truly progress. Then they added dungeon finder and for all I know the other four players could be bots for all the engagement you get.