I think you’re right, although it’s still not uncommon to encounter the ‘just an elaborate autocomplete’ argument. I believe what we’re seeing is that some level of reasoning is an emergent function of the ability to use the language coherently - that is, some of the capacity of human thinking is built into our languages.
Robert Miles described chat LLMs as (I’m paraphrasing from memory): a programming environment where the programming language is English and where the program and data are one and the same. It’s like an implementation of how language is used to perform cognitive processes, but ported onto a different architecture.