The startup Goodfire released Silico, a tool allowing engineers to adjust model parameters during training. This mechanistic interpretability approach enables direct modification of internal settings to control behavior. It offers more granular oversight than standard fine-tuning. Practitioners can now debug specific model failures by isolating and altering the underlying neural activations.