What is Goodfire?

Goodfire is a tool or method that analyzes model internals, as described by researchers. The material tracks its use in identifying unexpected dependencies in model predictions and improving sparse autoencoder (SAE) labels.

• Mar 2026 - Dan Balsam said Goodfire revealed that a model’s Alzheimer’s predictions depended overwhelmingly on fragment length, contrary to expectations from literature.
• Mar 2026 - Dan Balsam said Goodfire allowed model editing with no degradation in capabilities, with changes within noise levels.
• Mar 2026 - Geoffrey Irving said Goodfire adds another wrinkle to the mess but does not change fundamental aspects of model interpretability.
• Apr 2026 - Cameron Berg said Goodfire allows bootstrapping SAE labels by having the model label its own activations, yielding more accurate labels.