Maybe LLMs should follow best practices for 1980s style backprop models and later deep learning models: starve model size to force maximum generalization, minimal remembering.