I'm still not convinced that large language models are anything more than a fancy form of autocomplete, we solved the "generate human-like text" problem years ago with much simpler tech.