Nathan Lambert hypothesizes that some aspect of mid-training or common pre-training data makes it difficult to improve open models by simply fine-tuning on reasoning traces from top models