“According to Geoffrey Hinton, instructing a powerful AI to defend itself against attacks from an enemy AI would effectively give the friendly AI a survival instinct.”