Given the paper you must really understand AI. I feel like I have a good grasp of the math behind LLMs, but I don't understand them in that I don't understand

Given the paper you must really understand AI. I feel like I have a good grasp of the math behind LLMs, but I don't understand them in that I don't understand why people find them useful. I was part of a trial using them at work. The chat based ones were worthless for any problem harder than a CS400 class which is what we're actually paid to do. The autocomplete ones produces code so poor it was constantly yanking me out of flow.