Thoughts
We tried to do spell checking by giving the spelling checking model more low-level information about the relative position of keys
(especially on mobile, fancy mobile keyboards try to guess where your finger is supposed to be physically). But these models are absolutely blown away by ChatGPT, which doesn't have any low-level information about the position of keys at all, rather, it has tons of high-level information about the types of things that you want to say. I think this represents a kind of fundamental flaw in current/historical approaches to AI and real-world modeling.