Goodfire

Goodfire is a San Francisco–headquartered AI interpretability research lab (a public benefit corporation) that builds tools meant to help teams understand, monitor, and intentionally steer AI model behavior by working with model internals rather than treating models as black boxes. The company’s product is centered on “Silico,” an interpretability agent and infrastructure intended to turn research questions into repeatable experiments, reports, and (for some workflows) deployed guardrails. Silico is positioned for both (1) interpretability research workflows—such as activation harvesting, training and evaluating sparse dictionaries/featurizers, probing/monitoring, and automated interpretation—and (2) applied safety/monitoring workflows, including probe-based guardrails and related monitoring/cascades. Goodfire also publishes mechanistic interpretability research and technical tooling, including methods aimed at steering and training interventions, and has highlighted security, auditability, and reliability requirements for deploying interpretability-based techniques in production settings. In fundraising, Goodfire has progressed from a $7M seed round (led by Lightspeed Venture Partners) to a $50M Series A (led by Menlo Ventures) and a $150M Series B (led by B Capital) at a disclosed $1.25B valuation, and it reports having raised over $200M in total. Recent company updates include launching “Goodfire Research Grants” for academic and nonprofit researchers and announcing SOC 2 Type II compliance. On the customer side, Goodfire has publicly described use cases such as applying its Ember platform to build PII guardrails for Rakuten’s agent platform.