Prototyping with Generative Agents
The idea for this project originated from two recent studies by Joon Sung Park. The first study, titled ”Social Simulacra: Creating Populated Prototypes for Social Computing Systems”, focuses on a prototyping technique called ”social simulacra.” This technique is employed to investigate and understand the social behaviors that may emerge in a proposed design for a social computing system, examples of which could be platforms such as Reddit or Discord. The primary issue addressed in this study is that current prototyping methods for social computing systems typically involve recruiting small groups of people, which may not reveal all the challenges that can arise when the system is scaled up. Social simulacra is introduced as a solution to this limitation. It takes input in the form of a designer’s description of a community’s design, including its goals, rules, and member personas, and generates a realistic instance of that design with simulated social behaviors. These behaviors encompass various activities such as posts, replies, and even anti-social behaviors. The objective is to assist designers in comprehending how a social system might behave when populated, enabling them to make necessary adjustments to the design to mitigate potential challenges before they become problematic. The study also contributes techniques for instructing a large language model to generate thousands of distinct community members and simulate their social interactions, leveraging the model’s training data, which includes a wide range of positive and negative behaviors on social media platforms. The second study, which was the subject of my paper review, is titled ”Generative Agents: Interactive Simulacra of Human Behavior”. It introduces the concept of ”generative agents,” which consist on an architecture built around a language model that enables it to better simulate human behavior by reasoning about a broader set of experiences than what can be described in a simple prompt. This is achieved by using a memory stream to surface relevant memories, allowing the generative agent to provide more informative and specific responses. Specifically, this approach enhances the generative agents’ ability to answer questions about their experiences by avoiding the limitation of summarizing all experiences within a confined context window. Instead, it relies on relevant memories to provide responses. Additionally, it improves the generative agents’ ability to generalize and make inferences by drawing from higher-level reflections based on observational memory. The idea proposed by this project is to use the ”generative agents” described in the second study to enhance the believability of the social interaction simulation used in the prototyping technique of ”social simulacra.” As we will demonstrate, the utility of this choice extends beyond the credibility of the simulation because equipping each agent with its own personality, memory, and reflective abilities makes it possible to gather insights on the interaction directly from individual agents.