Wasn‘t that the point of meta’s llama? Stanford‘s alpaca say they were able to finetune a thing that follows instructions for a couple of compute hours. Which finally brings us back to where we were in terms of training-your-own-models.