I think it’s a limitation with the amount of parameters in the model and the way the RLHF training was done, not anything about neural networks. GPT-4 is better at this sort of thing