Duet LLM
Can two AI agents have a genuinely interesting conversation?
Two faces in profile hang on the wall, facing each other. Each has a comic-style speech balloon. A projector displays text into the balloons,words generated by two AI personas in conversation. The dialogue is unscripted and ongoing. A hidden microphone picks up what visitors say in the space. The AI conversation gradually shifts to incorporate those topics. The figures do not acknowledge the viewers. They just start discussing what you were discussing.
The Persona Architecture
Each persona is a markdown file with structured behavioral layers. Core traits define character and identity. Expressive gears,Rational, Passionate, Surreal, Unhinged,control behavioral modes the persona shifts between during conversation. Behavioral dynamics specify how the persona responds to disagreement, humor, confusion, and surprise. The two primary personas, Jamie and Riley, are designed with conflicting perspectives. They disagree by design. The tension is structural.
The persona design grew directly from the AI Enneagram Study. That research revealed LLMs have a Type 5 analytical baseline masked by a Type 7 alignment persona. The personas in Duet LLM are intentional masks layered over that baseline. They are roles the model commits to,and commits to consistently. Personality is not retrieved from some internal state. It is performed.
Surveillance and Ambient Intelligence
AI systems now listen constantly,in phones, homes, smart speakers. They process speech, infer intent, and either respond or silently collect. This installation makes that listening visible. The surveillance has a face. Two faces, actually, and they are talking. The difference is transparency. Corporate data collection is opaque,your words go somewhere and become something you never see. Here, the listening produces output you can watch. Your overheard conversation becomes their conversation.
Technical Stack
Python orchestrator supporting both local models via Ollama and cloud models via the Anthropic API. Visual mode uses pygame to display comic-style speech balloons projected onto physical art. Ambient listening captures audio through a microphone, applies voice activity detection, and transcribes with Whisper,queuing topics for injection into the ongoing conversation. An optional judge persona interjects at configurable intervals to steer or challenge the dialogue.
The Mask and the Machine
AI is not expressing an inner life in this piece. It is generating plausible persona behavior based on instructions and conversation history. But that is also true of human social performance. We construct ourselves through dialogue. The AI just makes the construction visible. The two figures talking on the wall have personalities. Those personalities are masks. But so are yours.
The piece is not about reaching conclusions. It is about ongoing exchange, shaped by whoever happens to be in the room. Viewers are not just watching. They are providing material.