Within a couple of years, LLMs or equivalent AI engines will have the ability to generate factual knowledge themselves using their "outdated" knowledge and their reasoning mechanisms.

For example, one of the classical requests from S.O. questions and GitHub issues is a "minimal reproducible example " of the question/problem.

So a sufficiently advanced AI will be able to write that, run it, see the issue , go to the library/related-system code or documentation (for closed source) and derive a solution