As models become smarter than their human trainers, evaluating whether their outputs are actually safe becomes computationally and cognitively impossible.
Open the full topic