Goodfire 发布 Silico 工具用于大模型机械可解释性
goodfire 發布了 silico，這是一個機械可解釋性工具，允許開發者映射和調整大型語言模型內部的神經元。該初創公司旨在通過暴露內部參數，使 ai 訓練更像傳統工程而非煉金術。
San Francisco startup Goodfire released Silico, a mechanistic interpretability tool that allows researchers to inspect and modify internal model parameters during training. The tool maps neurons and pathways to help developers debug behaviors and steer outputs, moving AI development closer to traditional software engineering.