AI Center Seminar - AI Fundamentals series - Simon Schrodi - "Towards a Science of AI (Safety)"

The talk is jointly organized by the EPFL AI Center and the DLAB as part of the AI fundamentals seminar series.

Host: Raghav Singhal

Title
Towards a Science of AI (Safety)

Abstract
AI models have remarkable capabilities, yet their successes and failures often remain poorly understood. More concerningly, what they learn can differ from what we intended. Understanding these gaps matters for assessing their reliability and safety. In this talk, I show how controlled experimentation and mechanistic analysis can help explain how models learn and generalize. Using examples from my research, I show that certain data properties but also internal mechanisms shape generalization, and how models can learn behaviors not explicitly expressed in their fine-tuning data. Together, these examples are steps towards a science of AI that helps us understand when, how, and why model behave as they do, providing us with a stronger foundation for their safe deployment.

Bio
Simon Schrodi is a fifth-year PhD student at the University of Freiburg, advised by Prof. Thomas Brox. His research lies at the intersection of the science of deep learning, interpretability, and AI safety. In particular, he investigates how AI models learn and generalize, and how this shapes their behavior using controlled experiments and mechanistic analysis. His broader interests also include machine learning for climate science and automated machine learning. Simon is also a research scholar at MATS 9, mentored by Alex Cloud, Cem Anil, and Arthur Conmy. His work there examines the limitations of pretraining safety methods and how pretraining shapes models' propensity for agentic misalignment. Previously, he completed his MSc in computer science at the University of Freiburg in 2022, following BEng studies at the Cooperative State University Karlsruhe. More information can be found at https://simonschrodi.github.io.