“The amount of data used for training state-of-the-art neural networks increased by a factor of approximately one million between AlexNet and GPT-4.”