“Achieving AI alignment is difficult, as suggested by over two decades of research in AI safety.”