When Lex Fridman interviewed Jeremy on his podcast, he mentioned that the latest research shows that you don’t need to tune training parameters but can actually choose them programmatically. Does anyone have any citations for this or know where I can read about this?