“Fully reverse-engineering superhuman AI systems via mechanistic interpretability is likely an intractable problem.”