"Human beings have been prompt injectable since their original development." @eddylazzarin reframes social engineering, training, and multisigs as the AI safety problems they actually are: "We use all types of human feedback mechanisms to refine humans and to make them less vulnerable to obvious types of social engineering attacks. But it's hard to do because human beings are so non-deterministic and so unpredictable and hard to measure and hard to instrument." "So they are very challenging to put the core loop of a system, which is why you have to put multiple of them, because it turns out we've discovered that adding multiple humans into a cybernetic process and requiring consensus among those human beings in sort of an ensemble setup is less prone to catastrophic error."(a16z crypto)