Goodfire AI

Goodfire AI is an AI interpretability research lab and public benefit corporation based in San Francisco. Its work uses model internals to detect unwanted behavior, extract scientific knowledge, and guide model training. Concrete applications include PII guardrails developed with Rakuten and reinforcement learning supervised by internal model features. On February 5, 2026, Goodfire announced a $150 million Series B at a $1.25 billion valuation to expand research, develop its next product generation, and build a “model design environment.”[5][8][9]

Those applications followed earlier work on Ember, Goodfire’s interpretability platform, and customer collaborations that began in its founding year.[3][4][8]

Goodfire was founded in 2024 by Eric Ho, Daniel Balsam, and Tom McGrath. Lightspeed’s portfolio record lists Ho as CEO, Balsam as CTO, and McGrath as chief scientist, and records a seed-stage investment in 2024.[3]

On April 17, 2025, Goodfire announced a $50 million Series A led by Menlo Ventures, with participation from Lightspeed Venture Partners, Anthropic, B Capital, Work-Bench, Wing, South Park Commons, and others. The company said the funding would expand research and develop Ember with customers. Ember was intended to give users programmable access to model internals so they could investigate learned concepts and shape behavior beyond inputs and outputs.[4]

Arc Institute was an early collaborator. In the Series A announcement, Arc co-founder Patrick Hsu said Goodfire’s tools helped extract biological concepts from Arc’s Evo 2 DNA foundation model: “Partnering with Goodfire has been instrumental in unlocking deeper insights from Evo 2, our DNA foundation model.”[4]

Goodfire and Rakuten began collaborating in 2024. Technical staff from both organizations used Ember to develop and test PII detection methods, producing a guardrail deployed to Rakuten’s agent platform. The system uses internal model signals to classify tokens containing personally identifiable information—such as names, addresses, phone numbers, and emails—before downstream processing.[8]

The deployment required token-level classification, allowing private information to be precisely removed rather than merely flagging an entire sentence. In the February 6, 2026 Latent Space interview, technical staff member Mark Bissell described English and Japanese requirements, including Japanese tokenization problems. Real customer PII could be used for evaluation but excluded from training, so methods trained on synthetic data had to transfer to real production data.[1]

Goodfire reported that sparse autoencoder (SAE) probes worked best for Rakuten’s synthetic-to-real transfer, non-English inputs, and noisy data. This was a task-dependent result: head of product Myra Deng said that in many other detection tasks, classifiers trained on raw model activations outperformed classifiers trained on SAE features. She attributed the Rakuten result to SAEs extracting more generalizable signals from noisier data.[8][1]

Customer work also exposed limitations in Goodfire’s control tools. Deng recalled that “we tried steering with our first API and realized that it still fell short of black box techniques like prompting or fine tuning.” That led the team back to research on more powerful, robust interventions. She credited Ekdeep Singh Lubana and Atticus Geiger with steering expertise and described customer problems as inputs to the research agenda, including work on methods beyond SAEs.[1][2]

In the same interview, Bissell demonstrated real-time activation steering on Kimi K2, which he described as a one-trillion-parameter model. Activating a feature associated with Gen-Z slang changed the model’s language while it continued calling tools. Bissell characterized the setup as hastily assembled and fragile; the demonstration illustrated frontier-scale engineering rather than production readiness.[1]

Features as Rewards, first submitted on February 10, 2026, extended interpretability from monitoring and steering into training supervision. Its authors—Aaditya Vikram Prasad, Connor Watts, Jack Merullo, Dhruvil Gala, Owen Lewis, Thomas McGrath, and Ekdeep Singh Lubana—introduced RLFR, Reinforcement Learning from Feature Rewards. The pipeline uses model features as reward functions and a probing framework to identify candidate hallucinated claims, teaching the model to correct completions when uncertain of their factuality.[9]

The authors reported that their end-to-end process on Gemma-3-12B-IT produced a policy 58% less likely to hallucinate than the original model when run with their probing harness, while preserving performance on standard benchmarks. The result applies to that combined model-and-harness setup.[9]

Goodfire also pursued interpretability for scientific discovery. Its February 2026 release named Mayo Clinic, Arc Institute, and Prima Mente as partners and reported identifying a novel class of Alzheimer’s biomarkers by interpreting an epigenetic model built by Prima Mente. Bissell described scientific work as including both debugging spurious correlations and extracting knowledge learned by foundation models.[5][1]

The February 5, 2026 Series B announcement named B Capital as lead investor. Participants included existing investors Juniper Ventures, Menlo Ventures, Lightspeed Venture Partners, South Park Commons, and Wing Venture Capital, alongside new investors DFJ Growth, Salesforce Ventures, Eric Schmidt, and others. Goodfire described aggregate backing of over $200 million and planned an environment for identifying model subunits responsible for behavior and targeting them with training or interventions.[5]

In the February 6 interview, Bissell said he had joined among approximately the first ten employees and that the team had grown to above 40. His work spanned image-model demonstrations, life-science partnerships, and the core platform. Deng described product work as translating research into repeatable deployments across engineering and research teams.[1]

By August 20, 2026, Goodfire was offering Silico, an interpretability agent for long-horizon experiments used by its research team and partners. The company described workflows including training probes and SAEs, exploring neural geometry, applying causal interpretability methods, and comparing models. As of September 18, 2026, Goodfire offered Silico for macOS and presented applications for language models, life sciences, robotics, and vision.[7][6]

On August 20, Goodfire announced research grants offering $1 million in free Silico usage to selected academic and nonprofit researchers working on interpretability, alignment, and life sciences. The announcement set out two research priorities: interpretability foundations and safety applications, including methods to identify and prevent reward hacking, cybersecurity risks, and biosecurity risks through model internals.[7]