“The ratio of compute spent on post-training relative to pre-training is increasing for advanced AI models.”