“A significant amount of work in major AI labs is dedicated to maintaining the stability of the deep learning training process, which is described as fundamentally fragile and prone to failure.”