AI reasoning models can think. Agentic models can act. But according to researcher Rafael Rafailov, today's most powerful AI still has a glaring blind spot — and fixing it might be the last thing we need to do before reaching superintelligence.