Class Central is learner-supported. When you buy through links on our site, we may earn an affiliate commission.

Generative Agents: Interactive Simulacra of Human Behavior

Center for Language & Speech Processing(CLSP), JHU via YouTube

Start learning Write review

Explore generative agents as interactive simulacra of human behavior in this Stanford lecture by Joon Sung Park. Delve into the concept of Smallville, a custom-built game world featuring common small village affordances, where agents interact with their environment and each other through actions and natural language dialogue. Learn about the memory stream storing agent experiences, the retrieval function for selecting relevant memories, and the generation of reflection trees. Discover how plans describing future actions are created using large language models. Examine the critical roles of observation, planning, and reflection in creating believable agent behavior. Consider the impact of instruction tuning on agent politeness and cooperation. Investigate the GENERATE system, which takes social designs as input to produce populated communities.

Syllabus

Intro
The ability to simulate believable human behavior promises a new class of interactive applications
A new opportunity: generative models trained today encode the way we live, talk, and behave
Setting: Smallville is a custom-built game wo featuring the common affordances of a small village
The agents interact with their environment through their actions
The agents interact with each other through natural Language dialogue
The users can interact with the agents via 1 dialogue, 2 altering the environment, and 3 embodiment
Memory stream stores a comprehensive record of agent experience in natural language
We retrieve a select portion of the agents' experien using a retrieval function
Over time, agents generate trees of reflections: the nodes as observations, and the non-leaf nodes as thoughts that become higher-level higher up the tree they are.
Plans describe a future sequence of actions for the agent that are stored in the memory stream. They include a location, a starting time, and a duration.
To generate plans, we prompt a large language moder with a prompt that summarizes the agent and the agent's current status.
Observation, plan, and reflection each contribute criticall the believability of agent behavior.
[Boundaries and Errors]: Instruction tuning seems to guide the behavior of the agents to be more polite and cooperative overall
GENERATE takes a social design as input and outputs community that might populate it.