If interpretability research goes well, it may get dangerous