LLMs can't consistently win at chess https://www.nicowesterdale.com/blog/why-llms-cant-play-chess

Now, some of the best chess engines in the world are Neural Networks, but general purpose LLMs are consistently bad at chess.

As far as "LLM's don't have understanding", that is axiomatically true by the nature of how they're implemented. A bunch of matrix multiplies resulting in a high-dimensional array of tokens does not think; this has been written about extensively. They are really good for generating language that looks plausible; some of that plausable-looking language happens to be true.