“The core technical challenge for AGI alignment is scalability, as current techniques relying on human supervision will fail when models become superhuman and their outputs are too complex for humans to evaluate.”