Chatter Generation through Language Models.

This work examines the feasibility of using Language Models (LMs) to generate chatter that stays in context based on persona descriptions. We clearly distinguish between chatter and dialogue, explain why we believe that chatter yields more promise for integration, and experimentally show that in 500 generated samples the majority (79%) of responses stayed in context. Additionally, we coarsely check that most (≈70%) consumer gaming hardware has enough random access memory (RAM) to store a small 7B 4-bit quantized LM model such as LLama.cpp on top of a demanding AAA game. Finally, we outline our vision for the future of games and language models and their potential synergies.