Understanding AI Alignment Through Learning Theory | Einar Urdshals | EAGxNordics 2025
Description
Join Einar Urdshals as he introduces how Singular Learning Theory (SLT) can help advance AI safety. As AI systems grow more powerful, we need to understand how they learn and generalize to ensure they remain aligned. Einar shares how Timaeus applies mathematical frameworks to connect training data, model structure and behavior. Discover why "you are what you eat" applies to AI systems, and how understanding learning dynamics could be key to building AI that reliably acts according to human values.Einar Urdshals is a Researcher at Timaeus where he applies Singular Learning Theory to AI safety challenges. With a background in theoretical physics and mechanistic and developmental interpretability, his recent work focuses on preventing weight exfiltration by studying theoretical limits of model compression.
United States
