AI Charades: Language Models as Interactive Game Environments

Abstract

Language models in recent years have shown astounding growth in modeling future text prediction. From a games perspective, language models present richly semantic environments, in which players can deeply influence the model's behavior through text input. This work presents a framework for building Language Model Games, games which center around a player manipulating a language model into behaving in a desired manner. A small demo, AI Charades, showcases a proof-of-concept game in which a player must communicate a secret word to an AI without mentioning it directly. Later discussion explores language models as a general tool for human-AI interfacing, along with their capabilities as design tools.

Topics & Concepts