What is Goodfire?

Goodfire is an AI interpretability tool or company. As of March 2026, Dan Balsam reported that Goodfire’s analysis revealed a model’s Alzheimer’s predictions depended heavily on fragment length—an unexpected finding not anticipated by the existing literature. Balsam also noted that Goodfire’s interventions showed “essentially no degradation” in model capabilities, with any observed changes falling within noise levels. In April 2026, Cameron Berg stated that Goodfire “allows you to bootstrap SAE labels, so that you can just have way more accurate labels on your SAE given basically having the model label its own activations.” As of March 2026, Geoffrey Irving offered a more cautious view, commenting that “nothing about that Goodfire thing changes that at all. It just adds another wrinkle to the mess.”