- The Regression Games Weekly Observation
- Posts
- LLM Agents Play Games and New AI Game Communities
LLM Agents Play Games and New AI Game Communities
Learn more about how LLMs are being used to play games
Aaron Vontell
May 16, 2023
Welcome to our weekly newsletter and platform update! Every week we’ll bring you the latest AI + gaming news, resources, and updates from the Regression Games team. A huge thank you to our 230+ readers so far! Share this newsletter with friends: https://regressiongg.beehiiv.com/
Weekly Spotlight - Agents Learn to Party
PC: J.S. Park, J.C. O’Brien, C.J. Cai, M. Morris, P. Liang, M.S. Bernstein
Researchers from Stanford and Google ran an interesting experiment where 25 virtual agents lived and socialized in a virtual world together. This miniature game focused on conversation and social game mechanics, and after seeding each individual agent with a separate personality, agents would chat, collaborate, and execute tasks. What really caught our eye was the emergent behavior that these agents demonstrated, or in other words, what complex behaviors emerged from the agents that they were not originally programmed to do. For example, asking an agent to plan a party resulted in proper coordination for that party.
You can read more about this exciting work through this Ars Technica article.
Blog Post - RL vs Imitation Learning?
This blog post from the Berkeley AI Research lab discusses the basics of offline RL and imitation learning approaches and compares them in different scenarios. If you’ve been curious about how imitation learning differs from typical offline reinforcement learning approaches, check it out!
Community - Hugging Face #ml-4-game-dev
There are many great communities out there where folks are exploring the limits of what AIs can do in games. The Hugging Face #ml-4-game-dev Discord channel is one of our favorites. Game developers and researchers explore and share ideas for getting AIs working in games (such as FPS bots and tower-building bots). We recommend joining it and checking it out by joining their Discord and selecting the ML for Game Development option.
Technical Paper of the Week
Every week we will select a recently-published paper for our reading club. This week’s reading: Can Large Language Models Play Text Games Well? What we love about this paper is that it’s extremely easy these days to run LLMs and text games (compared to something like a 3D game and RL models). With the integration of these two concepts being a breeze, anyone can experiment with this idea. Read the paper and run some of your own experiments!
Members of our community discuss these papers on our Discord #reading-club channel. You can also find all past resources on our GitHub.
Looking for something new? Join the Regression Games team!
Have feedback about our newsletter? Let us know at [email protected]
You are receiving this email because you registered to our mailing list or are a player on our platform.