Learning theory for AI safety - Guillaume Corlouer

Link to slides: https://docs.google.com/presentation/... Summary: Following Ari Brill's talk on physics-informed AI safety (   • Physics-Informed Research for Ambitious AI...  ), this session broadens the lens to the wider landscape of theoretical directions in AI safety, including where to start if you'd like to work on them. ​On Wednesday, July 8th, ‪@SafeAIGermany‬ hosted Guillaume Corlouer (Stormglass, ex Pivotal Research and ‪@PrincInt‬ ), to discuss: 1. Why understanding generalisation is important for AI safety, with empirical examples such as emergent misalignment, where undesirable behaviour emerges because of generalisation. ​2. Using toy models of deep neural networks for AI safety. Toy models are useful because they are mathematically tractable and can give us concrete hypotheses that we can test to understand specific safety-relevant phenomena. As a concrete example, we looked at deep linear networks, and how they can be used to understand the effects of depth and initialisation in deep learning. ​3. Some fellowships and organisations you can join, if you are interested in working on more theoretical directions in AI safety. ​Speaker profile ​​Guillaume Corlouer is a research scientist at Stormglass (formerly Pivotal Research and PIBBSS) and is working on developing toy models of deep learning that can be leveraged to improve our understanding of safety-relevant phenomena. ​Since 2023, Guillaume has been working in AI safety focusing on interpretability and better understanding the learning dynamics of deep neural networks. Previously, he did his PhD at the University of Sussex on estimating information flow in the brain, motivated by questions in the neuroscience of consciousness. About SAIGE: Safe AI Germany (SAIGE) is building Germany's infrastructure for safer AI. Our online events are open to everyone everywhere. 🌐 safeaigermany.org 📅 Upcoming events: luma.com/saige 🐦 Follow us on LinkedIn: linkedin.com/company/safe-ai-germany Recorded 8 July 2026.