“Safety” of own source code is actually a weak form of the problem. An AI has to keep the external world sufficiently “safe” as well, because the external world might itself host AIs or other dangers (to the external world, but also to AI’s own safety), that must either remain weak, or share AI’s values, to keep AI’s internal “safety” relevant.
“Safety” of own source code is actually a weak form of the problem. An AI has to keep the external world sufficiently “safe” as well, because the external world might itself host AIs or other dangers (to the external world, but also to AI’s own safety), that must either remain weak, or share AI’s values, to keep AI’s internal “safety” relevant.