People claim that finetuning is good because no model can be 'that' general since gpt3 and at every generation it's becoming less true