He's already a few weeks behind. Bing's new AI can query Bing's search engine. So it hasn't "burned the library". It can reach out for new information.
That's going to be really important. Large language models can already summarize documents. How long before you can tell a system "Read this book. Then I will have some questions about it"? Once you have that, you can apply the base model to domain-specific problems.
I think you missed the argument about the two Roman Empire books. ChatGPT was clearly trained on them, or at least metadata about them, yet it couldn't produce an intuitive response about their relationship. And it couldn't even return the top search result that explained it.
Sure, these issues will be fixed eventually, but the point of the article is that it can't replace the thinking of a student to write a quality essay. It might be able to regurgitate search results which will need to be fact checked, so at best it's useful as a research tool.
BTW, Bing's AI is a really poor counterexample, as it can't even return search results reliably.
That's going to be really important. Large language models can already summarize documents. How long before you can tell a system "Read this book. Then I will have some questions about it"? Once you have that, you can apply the base model to domain-specific problems.