Training an AI has always been a little like alchemy: trial, error, and plenty of hoping. Goodfire wants to turn it into something closer to engineering.
The AI lab is launching Silico, a commercial platform that packages up mechanistic interpretability — the technique of mapping a model's internal neurons and their connections to understand its behavior — into a product. Goodfire claims Silico is the first of its kind to span the full pipeline, from building datasets to training models.
Silico lets you zero in on a single neuron or a cluster of neurons, then follow the upstream and downstream pathways that connect them to the model's outputs. The pitch: instead of training blind, you can see what's actually happening inside and directly steer behavior.