I believe the parent poster's point is: LLMs are more effective when they use more memory, meaning the less they are forced to compress the training data, the b

I believe the parent poster's point is: LLMs are more effective when they use more memory, meaning the less they are forced to compress the training data, the better they perform.