“The interpretability research company Goodfire developed a technique that uses an interpretability signal within the model training cycle.”