Goodfire AI's Silico Replicates Classic Interpretability Papers

GoodfireAI's official account reported that its agentic research platform Silico autonomously replicated three established causal interpretability papers—Interpretability in the Wild, ROME, and Mixing Mechanisms—without access to the authors' code. Each replication began from a single prompt and proceeded independently. The announcement was shared by academic Kosta Derpanis. The event highlights an early demonstration of agentic systems performing independent research replication tasks in the interpretability domain.