The limits that come from the mechanism
These are not temporary bugs. They follow from how the thing is built.
It does not know about today. Training finished on a particular date. Anything after that is absent unless the product hands it to the model as text. Attach a search tool and this changes, but then the answer is only as good as what search returned, and the model will not tell you when the search came back thin.
It does not remember you. The model is the same fixed set of numbers for every person on earth. When a product appears to remember your name or your preferences, the product stored a note and quietly pasted it into the text at the start of your next conversation. Memory is a feature built around the model, not a property of it. That is also why it can be deleted, and why it can carry a wrong thing about you forever.
It does not learn from your correction. You point out an error, it apologises, and this conversation improves, because your correction is now part of the text it is continuing. Close the tab and it is gone. The dials never moved. Nothing changed for the next person.
It cannot check itself. Asking whether it is sure produces writing about being sure. There is no inner state it is reporting on.
It has no access to anything. No files, no accounts, no cameras, no internet, unless a person explicitly connected those and told you. When it says it will look into it and get back to you, that is a sentence, not a plan.
It has nothing at stake. It will not be embarrassed, sued, sacked, or hungry. When a doctor or an accountant gives you an answer, part of what you are buying is that they carry the consequence. That part is simply absent.
The limit people are wrong about
There is a fashionable dismissal: it is just predicting the next word, so it cannot really do anything.
That is as mistaken as the opposite. Predicting text well, at this scale, turns out to require a great deal. These systems write working code for problems that are not in any textbook. They translate between language pairs their training barely covered. They find the flaw in an argument they have not seen. They hold a long document in mind and answer a question that needs three separate parts of it.
Whether that constitutes understanding is a live argument among people who know far more than either of us will settle here. The practical position is the useful one: the capability is real, the failures are real, and neither cancels the other.
The honest way to hold both
Use it where being wrong is cheap and being wrong is visible.
A first draft you will rewrite. An explanation of something you will then verify. A summary of a document you have in front of you. Code you are about to run and test. Twenty ideas of which eighteen are useless.
Be careful where being wrong is expensive and invisible. Medical, legal or financial decisions you will act on without checking. Anything with a specific number in it that you will repeat to someone else. Anything about a real named person. Anything you are not equipped to evaluate, which is the hardest case, because that is exactly when a fluent answer is most persuasive.
Before you move on