[State of MechInterp] SAEs in Production, Circuit Tracing, AI4Science, "Pragmatic" Interp — Goodfire

Understanding an image model’s internal representations can create new ways to control its output. AI research company Goodfire’s Paint With Ember let users paint and drag concepts on a canvas. These gestures changed internal numerical patterns associated with visual content at particular image locations, giving users spatial control over generation.

The researchers discuss this application in the interview as a practical use of interpretability research. The public demo has since been retired, but the research illustrates how identifying what a model represents can support an editing interface.