Isn't this exactly one of the problems that GPT4 has been aiming to solve? It'll be interesting to see if this is the tyranny of AI -- whether degradation of NLMs is an increasingly insurmountable barrier to good performance or whether it's something that can be massaged out with better verification.