Michael Pound - Why AI Lies, and do we care? - Joy of Coding 2024

Few innovations have taken over our lives quite like the recent surge in AI. Everyone from startups to experienced industry veterans are busy trying to incorporate large language models into their workflows. In academia, we’re all scrambling to work out if AI will fix all our problems, or simply let the students cheat! Using these models to obtain reliable results isn’t as easy as we think: They’re capable of impressive writing, but also of lying about almost everything. In this talk I’ll explain how these models work and how they are trained. We’ll look at how the way they are trained means they will frequently make up or embellish facts, and how it’s difficult to control when they will or won’t do this. For industry practitioners, we’ll talk about whether we even mind that this happens. Perhaps we can use these models with care, even if they sometimes misbehave.