Post-training is an art form. @christinahkim says shaping model behavior means sculpting character. Every reward is a tradeoff. GPT-5 was a reset: balancing helpfulness with honesty, warmth without sycophancy. The assistant shouldn’t just answer. It should feel right to use.