The original investigation of LLMs was attempting to get them to understand language but what they found was that once the model understood language it also somehow understood the concepts, events, and things the language was being used to communicate. That's not missing the point, it's entirely the point of the current interest in LLMs because it's incredibly useful to have a model that not only understands how to construct a sentence but can also do a fair amount of reasoning and actual work with the information in the sentence.

The current default version of ChatGPT is primed to use search to answer questions which is fine in some cases but I personally almost never use the multi-modal version because the "classic" ChatGPT is much better at explaining things from its training data than it is when it just regurgitates search results. Now that should tell you something about the utility of optimizing for information content rather than just a lot of language use examples.