Interactive Synthetic Debates
Inspired by the metaphor of the “devil and angel” on our shoulders, this project explores the use of Large Language Models (LLMs) to create AI agents that engage in debates from opposing perspectives. The goal is to examine whether dissent among AI agents can deepen our understanding of topics, uncover hidden biases, and provide an engaging, reflective way to access diverse information. By fostering disagreement, this exploration seeks to reveal deeper truths and challenge conventional thinking.
What if we could listen to the little devil and angel that sit on our shoulders?
Like many others, I have been very curious about the capabilities of Large Language Models (LLMs) since they became available. Inspired by “Generative Agents: Interactive Simulacra of Human Behavior” by Park et al. and similar work, I have been particularly interested in the possibility of creating multiple agents that would interact with each other to explore what we might learn from their exchanges. Following this, I started working on a web application that would let me explore these ideas.
From collaboration to dissent.
I initially experimented with the idea of creating agents that would interact towards a common goal. Similar to autoGEN, but with a focus on the conversation rather than the final output. While the initial results were interesting, the agents tended to be very agreeable with each other, and I could not achieve much of the bouncing, disagreement, and iteration found in a collaborative creative process.
I then decided to go in the opposite direction: agents that would argue and disagree about everything. Could this push LLMs beyond their apparent neutrality and agreeableness? Could this also display the underlying biases of LLMs? Could these debates be a good way to learn and reflect on an issue? Could it be an engaging way to access information? I recently came across a quote from Oscar Wilde: “When people agree with me, I always feel I must be wrong.” This quote points to the value of dissent in reaching deeper truths, and seemed relevant for this exercise.
Building the system.
I brought this concept to life in a responsive web application using Retool and OpenAI’s GPT-4, a multimodal LLM for text generation. When a user enters a topic, the LLM generates two fictional characters with distinct names, goals, backgrounds, and opposing views. If an image is provided, the LLM describes it textually it to enhance the context. The system then simulates a debate between these characters, using their profiles and the contextual information. Users can join the discussion or continue generating exchanges between the fictional characters.
