“The startup Good Fire is developing mechanistic interpretability for AI models to understand their internal workings, rather than just evaluating their outputs.”