whichever LLM GP is using, there's undoubtedly mutliple JS engines' source code in the training data, making it not really "from scratch".