There is an increasing amount of evidence that using AI to train other AI is a viable path forward. E.g. using LLMs to generate training data or tune RL policies