It will also combine correct things in incorrect ways, so even if everything it is trained on is correct it will still print incorrect things. We see it in thi

It will also combine correct things in incorrect ways, so even if everything it is trained on is correct it will still print incorrect things. We see it in this article, both the statement "Indentation-based scoping, similar to Python" and the code with bracers were probably correct in their original contexts, but combined they are no longer correct. LLMs will make such errors no matter what kind of data you train it on, there is no known way to avoid that class of errors.