Goodfire API

Goodfire's Ember API is the hosted platform providing direct, feature-level access to AI model internals for precise behavior steering and mechanistic interpretability.

Ember is the first hosted mechanistic interpretability API, giving engineers control over AI models like Llama 3.3 70B. We use state-of-the-art Sparse Autoencoders (SAEs) to expose a model's internal 'features,' moving beyond the black box. This direct access allows for powerful interventions: teams are preventing jailbreaks with conditional feature steering and building highly interpretable classifiers (e.g., 75% accuracy in financial sentiment analysis using just three semantic features). Leading organizations (Rakuten, Apollo Research) leverage Ember to accelerate alignment research and deploy safer, more reliable generative AI.

https://www.goodfire.ai/