From my experience their capabilities are extremely narrow and generally perform terribly when faced with issues outside comparatively narrow training data.