“Many findings from AI interpretability research are likely to remain relevant even if specific architectural components, like the attention mechanism, are replaced in future models.”