How many (currently unknown) problems for alignment on the same level as inner optimizers will we encounter in very advanced AI systems?