Can we discuss how it’s possible that the paid model (gpt4) got worse and the free one (gpt3.5) got better? Is it because the free one is being trained on a larger pool of users or what?