“In the near to medium future, AI systems could gain the ability to rewrite parts of their own neural networks, contingent on progress in interpretability research.”