Sessions

Reading Group 9: Model Forensics

Detecting concerning behaviour does not establish misalignment. The paper proposes a protocol for investigating what drove the behaviour.

Reading Group 8: Global Workspace (deep dive)

A follow-up to session 7: instead of re-reading the paper, we dug into the code and the Neuronpedia demo to see the Global Workspace idea in action on Gemma.

Reading Group 7: A Global Workspace in Language Models

Anthropic locates a 'J-space' inside Claude that behaves like a cognitive-science-style global workspace: a routing hub for deliberate thought that could double as a monitoring surface.

Reading Group 6: Steering Along Manifolds to Control Neural Networks

A natural follow-up to the geometric calculator session: once curved geometry shows up inside a model, can you steer along it? Manifold steering outperforms the usual linear interventions.

Reading Group 5: A Geometric Calculator Inside a Neural Network

A geometric, manifold-based view of model features that pushes back on the Linear Representation Hypothesis.

Reading Group 4: Towards Guaranteed Safe AI

Reading Group 3: Auditing Language Models for Hidden Objectives

Reading Group 2: The Assistant Axis

Locating the assistant persona as a direction in activation space, and what it takes to keep a model tethered to it under drift.

Reading Group 1: Emergent Misalignment

How narrow harmful fine-tuning (e.g. insecure code) can induce broad, cross-domain misalignment, and what that says about how aligned behaviour is structured.