People just keep underestimating transformers. Big mistake. The architecture is incredibly capable.

LLMs are capable of keeping track of consistent behaviors and beliefs, and they sure try. Are they perfect at it? Certainly not. They're pretty good at it though.