any of the models that they "align" are clearly active processes. They don't simply say "don't talk about nukes"; they actively process user input to detect iss

any of the models that they "align" are clearly active processes. They don't simply say "don't talk about nukes"; they actively process user input to detect issues, and return NOOP or whatever to the larger model. There's zero sense they'd ever give you the raw model; we already know anthropic's paranoia about the chinese using its distillation.