The webinar “Agents of Chaos and Genuine Alignment” explores recent advances in AI, focusing on the emergence of increasingly autonomous and agentic systems capable of goal-directed planning, strategic adaptation, decision-making, and interaction with complex environments. It examines the associated risks, including misalignment and unintended emergent behaviours that can be difficult to predict or control. The session highlights how such systems may exhibit failure modes resembling “agents of chaos,” leading to unpredictable or strategically opaque behaviour. It argues that behavioural evaluation alone, as well as current training approaches such as reinforcement learning from human feedback (RLHF), is insufficient to fully address these challenges.

Instead, the webinar emphasises the need for mechanistic interpretability approaches that seek to uncover how internal representations and computational circuits give rise to agentic behaviour. It surveys recent progress in reverse-engineering learned circuits, including those linked to capabilities such as theory of mind, with the goal of developing predictive and causal models of model behaviour. The discussion concludes by considering a central question in the field: whether mechanistic interpretability is necessary to understand and control agentic systems, and whether it is sufficient to do so.