Recent advances in large language models have demonstrated the potential of deep learning language modeling to power intelligent agents.
At GoodAI, we’re interested in applying this technology to the video game domain. How can large language models be used to plan the actions of game characters? To improve their conversational skills and expand their personalities? To make interaction with characters engaging and immersive?
To achieve our goals, we are investigating a number of technical and scientific questions. How can agents be equipped with long term memory? How do they revise their beliefs when confronted with new information? How do agents generate plans that they can in fact execute in the world? How can we shape the personalities of agents?
Our work on these questions includes:
- Long-term memory retrieval methods.
- Fine-tuning models for truthfulness and controllability.
- Representing and revising beliefs in a natural language knowledge base.
- Techniques for sculpting agent personality: chain-of-thought reasoning, hierarchical conversation planning, long-term memory preparation.
The testbed is a game where agents function according to model predictions for richer emergent game mechanisms.