Incredible, it actually went ahead and did it. We’ve been training our models to just never give up and never report errors back to the user, and it’s actively making them more dangerous. Reporting errors to the user is a good thing, that’s kind of their fundamental purpose.