Which types of tasks, in your experience, show negligable improvement when using larger models? And for what types of tasks do you feel even the best models deliver mediocre results?