Same! As we talk about in the article, the failures were less from raw model intelligence/ability than from challenges with timing and dynamic interfaces