The ChatGPT Navier-Stokes thing seems like a clear-cut case of an LLM "front-running" human beings.
The obvious, inevitable result of that is that eventually there will be no human work to front-run, because nobody's going to do work just to have credit stolen by software.
And front-running humans is all LLMs can do and all they will ever do. Because we will get to the point where humans aren't even qualified to evaluate what LLMs are putting out, at which point advancement stops entirely, because hallucination machines cannot evaluate their own output.
I think this is the thing people really don't get: there is no basis other than religion on which to base the theory that self-improving software is possible. None.
One doesn't even really need to understand technology to be able to figure this out. You just need to understand human nature. The most prevalent feature of humans, especially technically inclined humans, is
HUBRIS.