San Francisco-based startup Goodfire has released a new mechanistic interpretability tool called Silico that enables researchers and engineers to examine the inner workings of AI models and modify their parameters during training, according to MIT Technology Review.
The tool allows users to peer inside large language models and adjust the settings that determine model behavior while training is underway. According to MIT Technology Review, this capability could provide model developers with more fine-grained control over how their AI systems operate.
Mechanistic interpretability is an emerging field focused on understanding the internal mechanisms of AI models, moving beyond treating them as black boxes. Silico represents Goodfire’s entry into this space, offering practitioners hands-on tools to debug and refine model behavior at a granular level during the development process.