Episode Details

An in-depth exploration of "The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models" by Christina Lu and colleagues. This paper examines how language models maintain their helpful assistant persona and identifies a single geometric axis in activation space that controls persona behavior. The research has significant implications for AI safety and alignment, showing that persona can be manipulated through activation steering with 80-90% success rates. We discuss the methodology, findings, safety implications, and what this means for the future of AI alignment. Paper: https://arxiv.org/abs/2601.10387 This podcast is from Colin Davis (colin-davis.com) using Claude & Elevenlabs.