“In the future, post-training could become the majority of the cost for developing large language models.”