What the researchers bolted on is an architecture that enables the storage and recall of memories, as well as self-reflection and more. They call out early the paper that even standard ChatGPT is not quite capable of this. ChatGPT here is used to provide the natural language abilities.

The paper goes in-depth on the details of the architecture they built, which is fairly extensive and uses many module instances of ChatGPT for specific purposes.

I would say what's 'groundbreaking' is their architecture of a 'recursive reflection' loop that allows the agents to generate trees of reflection on prior experiences; a long-term memory; novel approaches like conversational interrogation of the agents.

This comment was posted to a different thread, which we merged into the current thread:

Stanford's Groundbreaking AI Study Simulates Authentic Human Behavior - https://news.ycombinator.com/item?id=35520236