I am a mathematician, musician, and hiker. Intuitively, these things feel deeply interconnected to me, though many people are surprised at how orthogonal these interests seemingly are. At the very least, they each fill me with joy through a combination of beauty, love, and wildness.
I am currently working as a Researcher in AI Safety at the Gradient Institute, and previously at Timaeus. My work combines technical research in interpretability and evaluation methodology with communication of AI safety issues for policy audiences. I have worked on research for the UK AI Safety Institute and the Australian Department of Industry, Science and Resources, and my technical research uses Singular Learning Theory and Developmental Interpretability to study how neural networks learn.
Recent Work
-
Risk Analysis Techniques for Governed LLM-based Multi-Agent Systems
Risk identification and assessment for multi-agent LLM systems in organisational settings.
-
Dynamics of Transient Structure in In-Context Linear Regression Transformers
How transformers transition from general to specialised solutions during training.
-
Loss Landscape Degeneracy and Stagewise Development in Transformers
Connecting loss landscape geometry with developmental stages in transformers.
-
You Are What You Eat — AI Alignment Requires Understanding How Data Shapes Structure and Generalisation
Understanding how data shapes model structure is essential for alignment.
-
Distilling Singular Learning Theory
Accessible introduction to SLT and phase transitions in neural networks.