Goodfire

Overview

Goodfire is a mechanistic interpretability company applying interpretability methods to biology AI models. Co-founded by Daniel Balsam (CTO) and backed by Menlo Ventures, Goodfire builds tools to map a model's internal representations to human-interpretable biological concepts — effectively an "AI debugger" for scientific models. The company's core thesis is that AI models should be designed like written software, with understood, inspectable internals, rather than grown organically as black boxes.

In a notable collaboration with the Mayo Clinic, Goodfire applied interpretability methods to EVO 2 and discovered that the model had learned to classify pathogenicity from ClinVar variants without being explicitly trained to do so — an emergent latent capability revealed only through mechanistic analysis.