The fact that the GPT-style models are literally autocomplete engines, while text autocompletion on phone keyboards remains terrible, really maddens me. Is it that they simply aren't using that style of model for autocompletion, or that the GPTs aren't great at guessing me specifically even though they're superhuman at autocompleting the Internet, or that it would be too slow to run a sufficiently good GPT for this on-device, or what?