[rsanek]

I'm not sure this is necessarily the case, especially considering techniques like RLHF that can improve a model's performance.