I think ChatGPT is pretty convincing when asked to generate an Airflow DAG that queries a date partition of a Google BigQuery table and checks if it has 24 unique values for a column in that partition.
It generates some code that by a glance can immediately put into Pycharm and execute without much modification. I then up the difficulty by asking for another task that fetches a secret from Google Bucket and use it to query an API and store the result into a BigQuery table. It surprisingly writes some good code that I can modify and use. Of course it doesn't know exactly know what data I need and which table to store, but the boilerplate part is pretty robust.
This leads to think that once we can feed ChatGPT with some conditions (for example feed it with some sample code in production), we might be able to develop rather quickly. It's not going to replace programmers for sure, but will accelerate a lot of mudane tasks -- and most tasks are indeed repetitive and mudane.
For example, I might ask it write a function and get one that's mostly right, except for one thing. I'll tell it 'great, just change_one_thing' and it will rewrite the function function differently (switching a structure from a tuple to a dictionary, or adding in calls to functions that don't exist etc.).
It helps a lot to start with an outline and use consistent labels in instructions. Also, to bear in mind that it will almost invariably charge ahead with its best guess rather than asking a clarifying question.