How so?

The models get better and better at understanding the intent of a prompt and doing more useful work with less intervention.