Goodfire is an AI interpretability research lab and public benefit corporation based in San Francisco, California. It builds interpretability agents and infrastructure intended to help researchers understand, monitor and align advanced AI models. Its work spans mechanistic interpretability, neural network reverse engineering, large language models and AI alignment, with applications in life sciences, healthcare and robotics.
The company's flagship product, Silico, decodes the neurons inside AI models to provide direct, programmable access to their internal workings. Goodfire reports that interpretability techniques built on this work enabled a 58% reduction in large language model hallucinations, and that reverse-engineering foundation models led to the identification of novel Alzheimer's biomarkers.
Goodfire has raised over $200 million in funding. Investors include Menlo Ventures, Lightspeed Venture Partners, Anthropic and B Capital. The company maintains collaborative partnerships with academic and industry organisations including Arc Institute, Mayo Clinic, Microsoft and Basecamp Research.
Founded by researchers who helped pioneer interpretability at OpenAI and Google DeepMind, Goodfire describes its mission as understanding and intentionally designing advanced AI systems. The lab is organised around fundamental science and the reverse engineering of neural networks.
