I use an LLM to do stuff that I find tedious and can easily verify to be correct (e.g. creating arguments for a script using argparse), or turning something from a table in a PDF into a Python list. My experiments trying to get any level of reliability for more complex tasks have been infuriating failures. They invent parameters and functionality that doesn't exist, swear blind that something is true but can't provide accurate references (or provide references that directly contradict what they just said), and so on.