Fields Mathematical AI Seminar

When AI agents meet and interact, their behavior is currently underspecified by training. There is a window of opportunity right now to seed this "multi-agent void" with examples of AI agents collaborating, with humans and with each other, in ways that humans endorse. These early examples of agent interactions can influence the training prior for the next generation, in the same way that early examples of the "helpful assistant" persona shaped the prior for today's chatbots. I'll outline a mathematical model of self-reinforcing priors, and describe a small ongoing experiment.