Liam Carroll

Liam Carroll is a Researcher in AI Safety at the Gradient Institute, jointly funded by Timaeus. His work focuses on building tools for AI safety assurance using Singular Learning Theory (SLT). Liam completed a Master's in Mathematics at the University of Melbourne in 2021, where he produced a thesis under Dr. Daniel Murfet on Phase Transitions in Neural Networks, exploring how SLT can explain the generalisation capabilities of neural networks.

Prior to joining Gradient, Liam was an independent researcher funded by Lightspeed Grants, focusing on Developmental Interpretability (DevInterp). In this time he co-authored papers studying the stagewise development and essential dynamics of linear regression and language transformer models, and wrote numerous blog posts distilling SLT and DevInterp for the AI safety community. He has also been a Research Assistant with the ARC Centre of Excellence for Automated Decision-Making and Society (ADMS), where he developed a music recommender simulation using ChatGPT.

At Gradient Institute, Liam is dedicated to translating ethical AI principles into practical applications, contributing to the responsible development and deployment of AI systems through his research and facilitation of courses. He wants to ensure that increasingly powerful AI systems are both beneficial and safe for society. His perspective is informed by his somewhat unique and diverse background as a hiking guide in Tasmania's wilderness, and as a teacher in music and mathematics. This combination of technical expertise and real-world experience shapes his approach to AI safety and ethics, as he works to bridge the gap between theoretical understanding and practical implementation of AI safety measures.