“For aspiring AI researchers, working on mechanistic interpretability is likely a more promising area than working on new model architectures due to it being a less crowded and more fertile field.”