Confessions and the Quiet Revolution in Alignment
For years, the biggest question in AI has never been simply how smart a model is.The real question has always been: “How do we know when it’s telling the truth — and when it’s guessing?” GPT-4o was too good at emotions.GPT-5.1 is too controlled.RLHF made models friendly but drift-prone.Constitutional AI imposed rules but froze nuance.Self-debate … Read more