Conversely, this also implies our current model sizes can still embed a ton more “understanding”