Maybe interpretability is free, maybe it isn’t.
One story of how it might be free is that it might get the model out of local optima and make global coordination in the model easier when there are modules within the model that have clear interfaces.
Maybe interpretability is free, maybe it isn’t.
One story of how it might be free is that it might get the model out of local optima and make global coordination in the model easier when there are modules within the model that have clear interfaces.