a popular measure for how well-secured the AIs’ utility functions were against self-hacking ended up being how many paperclips they produced before reaching nirvana