This isn't the reason. Models are pretty good at understanding relative positions. We put that in them and reward it a lot.

The issue is the same as why we don't use LLMs for image generation. Even though they can nominally do that.

Image generation seems to need some amount of ability to revise the output in place. And it needs a big picture view to make local decisions. It doesn't lend itself to outputting pixel by pixel or character by character.

Great point, but you need to have a good understanding in how LLMs work to understand this limitation. If you don't have an intuitive understanding think like it's one of these draw on my back games, just each new token is a new human in the loop, known words are like simple shapes you felt before 100 times on your back and are easy for you to reproduce and change, random ASCII strings are harder to _grasp_ and will produce a fuzzy output... all models are wrong, but some are useful.

https://www.youtube.com/watch?v=bA_DQHoxj34