Matthew Kowal - Interpretability Across the Model Design Pipeline

​Interpretability can be understood through the lens of model affordances: the actions a system makes possible. In this talk, Matthew Kowal, MTS at Goodfire, will argue that interpretability gives researchers and engineers new ways to intervene throughout the model design pipeline. The first part of the talk will introduce recent methods for creating better "knobs" to understand and control model behavior by decomposing model activations into multidimensional manifolds rather than straight lines. He will then discuss ongoing work on using these interpretable knobs during training itself, drawing on a model's own representations to shape its learning while keeping the process interpretable and monitorable.