TinyTroupe
An experimental Python library that simulates people with distinct personalities using LLMs, for testing products, generating synthetic data, and gathering persona-based feedback.
TinyTroupe is an experimental Python library that simulates people with specific personalities, interests and goals, called TinyPersons, who can listen to each other, reply, and act within simulated TinyWorld environments, using large language models such as GPT-4 to generate realistic behavior. Its focus is on understanding human behavior in a simulation rather than directly supporting it the way an AI assistant does.
Application ideas described include offline evaluation of digital ads with a simulated audience, providing test input to systems such as search engines or chatbots and evaluating the results, generating synthetic training or exploratory data, giving feedback on project or product proposals from the perspective of specific personas such as physicians or lawyers, and simulating focus groups for product feedback.
The project is released at an early stage, with a preprint paper describing the library and its use cases, and related experiments in its publications folder.