DeepMind’s claim that large language models were being trained with a suboptimal use of compute was also verified independently later by Google AI’s research.