By driving down the costs of training and inference, and then encouraging experiments. For LLMs, QLoRA is arguably a great step in this direction.