Goodfire

Goodfire is a research company and public benefit corporation based in San Francisco that applies mechanistic interpretability — the study of how neural networks work internally — to understanding, learning from and designing AI systems. The company positions itself against the prevailing "black box" approach to model development, arguing that direct access to internal model structures allows teams to steer what models learn, diagnose failures, and extract knowledge embedded in the models themselves. Its stated goal is to make AI that can be understood, debugged and shaped like written software.

The company's commercial product line has evolved from Ember, the model-agnostic interpretability platform described at the time of its Series A, which decoded neurons inside a model to give programmable access to internal representations, to Silico, described on the company website as an "interpretability agent" and platform for "intentional model design." Silico is offered for large language models, life sciences models, and robotics and vision models, is distributed as a macOS download and via team/infrastructure deployments, and is intended to let users see what a model has learned, locate undesired behavior, and make targeted interventions. Goodfire also describes building a broader "model design environment" for understanding, debugging and intentionally designing models at scale.

Alongside product work, Goodfire publishes a continuous stream of fundamental and applied research, including work on sparse autoencoders and neural geometry, steering along manifolds, interpreting language model parameters, memorization and loss curvature, evaluation awareness, probe-based data attribution, and applied studies with partner organizations in genomics, cardiology, materials discovery and robotics.

Founding story

Eric Ho previously founded RippleMatch, a Series B AI recruiting startup; Dan Balsam was founding engineer and head of AI there. Tom McGrath previously co-founded the interpretability team at DeepMind. The three founded Goodfire as a research lab devoted to studying the mathematical structure and hidden representations inside neural networks, on the premise that AI development is gated by fundamental science in the way earlier engineering disciplines were.

Business model

Goodfire develops interpretability software and research services for organizations that train and deploy AI models. Its website markets the Silico platform through demo requests and direct download, with a published pricing page, and it also engages in collaborative research and partnership arrangements with model developers and scientific institutions.

The company offers its interpretability platform commercially, with a pricing page and paid access to Silico for teams and their own infrastructure, alongside partnership engagements with model developers and research institutions.

Traction

Goodfire reported partnerships spanning scientific discovery and model design: interpreting Arc Institute's Evo 2 DNA foundation model, with findings published in Nature and used to explain 4.2 million genetic variants; identifying a novel class of Alzheimer's biomarkers from an epigenetic model built by Prima Mente, described as the first major natural-science finding obtained by reverse-engineering a foundation model; deploying SAE probes for PII detection in production with Rakuten; and analyses of a cardiac echocardiography vision model and a robotics policy. It raised $50M in April 2025 and $150M at a $1.25B valuation in February 2026, and obtained SOC 2 Type II certification in May 2026.

Latest developments

In February 2026 Goodfire announced a $150 million Series B at a $1.25 billion valuation led by B Capital, earmarked for frontier research, the next generation of its core product — a "model design environment" — and scaling partnerships across AI agents and life sciences. It subsequently announced SOC 2 Type II certification (May 2026), launched Silico as an interpretability agent available for macOS download, announced a research grants program (August 2026), and continued publishing research including work on neural geometry, uncertainty dynamics in text generation, and predictive data debugging.