Seriously, the «best thing they can do» is probably to "make it a module" and build the interfaces to integrate it in systems where it is controlled.

I use it for brainstorming ideas for writing. The fact that it can come up endless ideas incorporating whatever surface level details you demand makes it extremely effective for this purpose. At least so far I've not seen it produce quality you would want to use that wasn't filtered through a creative human (even the good humor and poetry it produces on its own is cherry picked), but we shouldn't underrate the value as a brainstorming tool. And the day is probably not too far away where it can filter for itself.

> even the good humor and poetry it produces on its own is cherry picked

This is so funny in itself, and speaks so much of human expectations :)

Yes, the really good examples are cherry picked, but there's a loooot of "good enough" examples and some cherry-picked bad examples of ChatGpt.

To me this is mind boggling. My ML knowledge stopped at 2018 levels and ChatGPT makes me think picking up ML once again.

I remember training a ML model in 2018 in a yandex course for image description. Up to the last few epochs the output was gliberish, then a few epochs later booom, readable english.

I've used it to help come up with metaphors in Japanese. Have to filter out its misuses of words, but still, very helpful source of ideas.