A team at Google DeepMind that studies something called mechanistic interpretability has been working on new ways to let us ...