San Francisco-based startup Goodfire released Silico, a tool that enables researchers to peer inside AI models and adjust parameters during training. The company aims to transform AI development from alchemy to precision engineering, addressing the widening gap between model capabilities and our understanding of how they work.

Silico focuses on mechanistic interpretability, allowing users to examine individual neurons or groups of neurons, trace pathways, and experiment with model behavior. The tool has already helped reduce hallucinations and identify specific neural circuits related to ethical reasoning in models like Qwen 3.

By packaging advanced interpretability techniques into a commercial product, Goodfire puts powerful debugging capabilities in the hands of smaller firms and research teams. The tool helps identify and adjust behaviors that might otherwise be opaque, potentially making AI systems more transparent and trustworthy for critical applications.