Goodfire is an AI interpretability research lab that builds tools to understand and design the internal mechanisms of neural networks. Their flagship product, Ember, gives engineers direct, programmable access to AI model internals.
Goodfire is an AI interpretability research lab that builds tools to understand and design the internal mechanisms of neural networks. Their flagship product, Ember, gives engineers direct, programmable access to AI model internals.
People
Updated 05/18/26Co-Founder & CTO
Co-Founder & CEO
Co-Founder & Chief Scientist
Board Member
Funding Details
Updated 05/18/26- Annual Budget
- -
- Current Runway
- -
- Funding Goal
- -
- Funding Raised to Date
- $207,000,000
Org Details
Updated 05/18/26Goodfire was founded in June 2024 in San Francisco, California, by Eric Ho (CEO), Dan Balsam (CTO), and Tom McGrath (Chief Scientist). Eric Ho previously founded RippleMatch, a Series B AI recruiting startup, and became convinced that AI safety was the most critical challenge of the era. Tom McGrath co-founded DeepMind's mechanistic interpretability team and left Google in early 2024 to pursue commercializing interpretability research. Dan Balsam served as RippleMatch's founding engineer and joined to lead the technical build-out. The company's core premise is that AI systems should not be black boxes. Goodfire's approach — mechanistic interpretability — involves decoding the individual neurons and internal features of neural networks to understand precisely how they represent concepts and generate outputs. This research tradition, pioneered at OpenAI and DeepMind, had largely remained in academic settings; Goodfire's founding thesis is that a well-capitalized commercial lab can scale this work faster and apply it more broadly. Goodfire's flagship product is Ember, a hosted mechanistic interpretability API and SDK. First shipped in December 2024, Ember supports models including Llama 3.3 70B and 3.1 8B, and enables engineers to identify model features, steer behavior, remove unwanted knowledge, and improve adversarial robustness. In May 2025, Goodfire released Paint with Ember, extending these capabilities to image generation models. The platform operates on a usage-based pricing model. The company has grown rapidly since its founding. A $7M seed round in August 2024 was led by Lightspeed Venture Partners. A $50M Series A in April 2025 was led by Menlo Ventures, with participation from Anthropic (marking Anthropic's first direct investment in another company), B Capital, Work-Bench, Wing, and South Park Commons. In February 2026, Goodfire closed a $150M Series B led by B Capital at a $1.25 billion valuation, with participation from DFJ Growth, Salesforce Ventures, Eric Schmidt, and existing investors. Total funding raised exceeds $200M. As of early 2026, Goodfire employs approximately 51 people. The team includes Nick Cammarata (a core contributor to OpenAI's interpretability team), Leon Bergen (professor at UC San Diego, on leave), and numerous PhD researchers from DeepMind, OpenAI, Harvard, MIT, and Stanford. Notable scientific collaborators include the Arc Institute (applying Ember to DNA foundation models) and Mayo Clinic (identifying novel Alzheimer's biomarkers by reverse-engineering a foundation model).
Theory of Change
Updated 05/18/26Goodfire believes that the primary bottleneck to safe and beneficial AI is not capability but comprehension: we cannot reliably align, audit, or improve systems we don't understand. By advancing mechanistic interpretability — decoding the neurons and internal features of neural networks — Goodfire aims to give researchers and engineers the tools to detect misalignment (sycophancy, deception, unfaithful reasoning), surgically edit unwanted model behaviors, and trace reasoning for high-stakes applications. The causal chain runs from interpretability research to practical tools (Ember), to widespread adoption by AI developers who use these tools to build safer and more reliable systems. A secondary path to impact is scientific discovery: by extracting insights from superhuman foundation models in domains like genomics and medicine, Goodfire's work could accelerate beneficial science while demonstrating that understanding AI internals is tractable at scale.
Grants Received– no grants recorded
Updated 05/18/26Projects
Updated 05/18/26Hosted API and SDK that let developers control a model’s internal “features” to inspect, edit, and shape AI model behavior beyond black-box inputs and outputs.
Mechanistic-interpretability tool that lets researchers inspect a model’s internals and adjust parameters during training to debug datasets and model behavior, aiming to turn AI development into precision engineering.
Discussion
No comments yet. Be the first to share your thoughts.