Goodfire (Batch 4)

Goodfire builds production tooling for mechanistic interpretability, the subfield of AI safety research focused on reverse-engineering internal computations of neural networks. The company was founded in 2024 and selected for AI Grant Batch 4 the same year, receiving a $250,000 SAFE plus Microsoft Azure and partner credits per AI Grant's terms. Its founding team includes researchers who authored cited mechanistic-interpretability work, with sparse autoencoders (SAEs) for feature discovery as a primary methodology.

In April 2025 Goodfire announced a $50 million Series A led by Menlo Ventures, with Anthropic participating; Anthropic's $1 million investment was its first equity check into another company per Dario Amodei's public statement. Other Series A investors include Lightspeed Venture Partners, B Capital, Work-Bench, Wing, and South Park Commons. The round valued the company at approximately $200 million per Contrary Research's company breakdown.

Goodfire's main product is Ember, an interpretability platform that trains sparse autoencoders on production LLMs and exposes the discovered features through an API and SDK. Customers use Ember to inspect, steer, and modify model behavior at the feature level rather than through prompt-level interventions. Goodfire has open-sourced SAEs for Llama 3.1 8B and Llama 3.3 70B (the Llama 3.3 70B SAE is trained on layer 50 with an L0 sparsity count of 121 features), published as the `Goodfire/Llama-3.3-70B-Instruct-SAE-l50` repo on Hugging Face. Their public research line includes "Understanding and Steering Llama 3 with Sparse Autoencoders" and an open-source SAE release announcement covering the Llama 3 line.

Within the open-source AI stack Goodfire sits at evaluation (interpretability as a kind of post-hoc capability and behavior evaluation) and at weights (their open-sourced SAEs are first-class artifacts that derive from but operate on open-weights base models). The company is one of a small number of mechanistic-interpretability vendors operating at production scale alongside in-house teams at Anthropic, OpenAI, and Google DeepMind.