I think the fundamental problem is that AI companies have been assuming that reinforcement learning with human feedback is an adequate foundational technology for guardrails. And that simply isn’t true.