Introduction

Sotopia is an AI tool whose main purpose is to conclude the intelligence of other AI tools and agents. It was created by a team at CMU led by Xuhui Zhou, a PhD student in the Language Technologies Institute and Hao Zhu, a former PhD student. Some other contributors to the project are Leena Mathur, Ruohong Zhang, Haofei Yu, Zhengyang Qi, Louis‑Philippe Morency, Yonatan Bisk, Daniel Fried, Graham Neubig, and Maarten Sap.

What is this AI tool?

The AI tool evaluates the social intelligence of other AI agents. The team believes that AI tools can do tremendous things but they are likely socially unaware or even biased and display harmful behaviors. Sotopia is an agent which is supposed to connect human and AI interaction stronger than before and situations in Sotopia typically public and secret information to determine how the AI agent will react.

How does this AI tool work?

Sotopia decides the AI agent’s social intelligence by simulating complex social interactions between the large language model’s artificial agents. The artificial agents and characters in Sotopia have names, a gender, a personality, job, public information and secret information. This allows for a lot of different scenarios and many different outcomes while testing AI tools and how they react.

CMU’s Impact

The development of Sotopia at CMU has made a huge impact because it highlights the fact that even extremely advanced models of today do worse at social interactions than humans. Sotopia also evaluates other variables like believability, relationship maintenance, sharing information and keeping secrets. In summary, the work of the team at CMU in developing Sotopia has made a major impact because of the ability to evaluate not only effectiveness of AI agents but also social variables.

By:
Aarush