What happens when you throw an AI into a group chat with real people playing Mafia - the classic social deduction game? I wanted to find out. But there was a twist: group chats are messy and *asynchronous*. Humans decide *when* to speak based on the flow of conversation, not just *what* to say. Most LLMs today are built to respond to direct prompts - but that's not how us humans play Mafia. So I built an LLM agent that decides both what to say and when to say it. I tested it in full games against real human players. It had to argue, deceive, stay quiet at the right moments, and almost half of the players couldn't even tell it was a bot. In this talk, I'll share the behind-the-scenes story: the agent's architecture, the crazy edge cases, the human players who got suspicious, and what it taught me about AI timing, social dynamics, and building agents that can survive the chaos of real group interaction. If you've ever wondered how far LLMs can go beyond chatbots - join me to find out!
Room: Main hall
Mon, Oct 27th, 16:20 - 16:50