“OpenAI uses a "mid-training" step to extend a model's intelligence and update its knowledge cutoff without the cost of a full new pre-training run.”