Who Gets to Build the Future?

Reinforcement Learning (RL) is a powerful branch of AI inspired by Pavlovian trial-and-error learning, capable of solving complex problems and often converging to "superhuman" performance levels. But there's a catch: Since RL learns through trial and error over many attempts, it often has to make catastrophic mistakes, many times, before it figures out which actions are undesirable and should be avoided. But... does it really have to be this way? Why should we let an AI explore in ways we already know to be suboptimal? What if we could build systems that simply never even consider ideas that are obviously bad for us?
In this talk, we'll explore how to design safer AI systems by integrating our existing knowledge, for example in the form of safety constraints, directly into the learning process. Instead of punishing bad behavior after it happens, we'll see how to build agents that are incapable of taking forbidden actions, even as they explore and learn. I'll begin with a brief introduction to RL, then walk through classical approaches to constrained learning, and finally talk about my contemporary research on tools like normalizing flows to embed constraints and human preferences into the AI's decision-making from the start.
Come find out how we're making AI not just intelligent, but safe, trustworthy, and aligned with human values -- without holding back its creativity.

Why do so many girls lose interest in tech before they even get started? And what can we do to change that? In this talk, Ulrika Sultan explores the hidden messages and cultural stereotypes that quietly push girls away from science, maths, and technology. Drawing on her research, she shows how early experiences shape who feels like they “belong” in STEM. You’ll hear stories of what works, what doesn’t, and how small shifts in everyday life can help all kids - especially girls -see themselves as future engineers, scientist, and creators.