The article is from 2020. Recent LLM developments show that a model trained on math and code is better at coding than one that's trained on language and code without math.