very exciting work from some of my colleagues at Goodfire. it turns out we *can* train on interp and make meaningfully positive changes to model behavior, as long as we understand how the model shifts in response (which we definitely need to be careful about!)
We cut through the mystery around interpretability in training.
A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities.
We call this *mechanistic shaping*, and we’ve used it to mitigate
