A question that comes up in my mind, since I don't fully understand how this works, is how does this affect runtime performance. It feels like abliterated weight models would work faster than some extra runtime operations?